Information Systems Institute, University of Salford, Salford, M5 4WT, U.K. A.Basden@salford.ac.uk
Copyright (c) Andrew Basden, 2001, all rights reserved.
Keywords: Knowledge level, Theory of levels, Philosophical underpinnings, Dooyeweerd, Presuppositions, Perspectives, Principle of Rationality, Representation.
Allen Newell's knowledge level, and indeed his whole theory of levels, has lasted twenty years and there is no reason why it should not last another twenty. It was published first in 1982, in a paper called The knowledge level [KL] and Newell made subsequent comments on the notion ten years later in [RKL] and [UTC]. In the companion to this paper, The knowledge level - the first twenty years, [CP], we examined Newell's notion of the knowledge level in some depth and surveyed how it had been used over the last twenty years. We noted a close parallel with other suites of levels, notably those found in linguistics, and from that suggested there is a level above the knowledge level, the tacit level, that adds a social dimension of cultural assumptions. We reviewed a surprisingly wide range of uses of Newell's ideas, many of which Newell himself had not envisaged, in particular moving outside artificial intelligence to human factors. The most important use was the least exciting, in some ways: to validate our intuition that knowledge can be treated in its own right, separately from symbols that represent it and not as a branch of psychology, sociology, etc. Newell had provided not only a label for this (the knowledge level) but also worked it out as a coherent framework, in some detail and made a bold ontological claim for it.
Now we prepare for the next twenty years. Newell's theory is not without problems, and it might benefit from some refinement to make it robust over time and in new contexts of use. If we are to refine an idea, then we should do so sensitively, respecting the original motivations, concerns and core proposals of its author. We should also do so in a coherent manner rather than just making ad-hoc changes from time to time. So we need to find a philosophical underpinning for Newell's ideas.
But surprisingly little discussion of Newell's theory has taken place. While there has been widespread use of, and reference to, it there has been very little critique and no comprehensive survey of the theory's problems. This is what we attempt in this paper.
In this paper we first summarize Newell's theory, outlining its key points (labelled 'kp'), for those who have not read the companion paper [CP] in which a fuller account exists, and to provide a reference point for use in the rest of the paper. The key points are listed in groups in appendix 1, and the reader may wish to refer to this while reading, though a fuller version is found in appendix 1 of [CP].
Then we examine a number of problems with the theory. Some are mere confusions, some have arisen because of new situations, but some are more fundamental. However, we will find there are surprisingly few major problems, and this motivates us to seek a philosophical basis for refining the theory.
If we try to address the problems individually and directly, the danger is that we will do so in a piecemeal fashion that robs the theory of its coherence, or that in response to a specific context we will jeopardise the generality of the theory while leaving underlying problems unsolved. Contradictory modifications might be proposed by different communities and the theory becomes split in two. To avoid such dangers we establish philosophical underpinnings of Newell's theory that shows how all the key points fit into an entire scheme, so that they do not become dislocated when the theory is modified. Providing such a foundation will show that Newell's theory is philosophically coherent - in fact remarkably so - and thus likely to be of long term relevance across a wide range of contexts, not just a useful device or a particular interpretation of one community for one short period of time. We will also see that conventional philosophies, whether of rationalist or post-rationalist tendency, cannot support Newell's whole theory.
Therefore, after noting the root of this difficulty in presuppositions made by most theoretical thinking, we outline one new stream of philosophical thinking that makes different presuppositions, the philosophical framework proposed by the Dutch philosopher, Herman Dooyeweerd (1895-1977). We explain what is needed of Dooyeweerd's ideas to support each of the key points of Newell's theory, finding a surprisingly good match, and briefly comment on ways in which the philosophy helps to solve its problems. This, we hope, will ensure Newell's theory remains relevant, powerful and useful over the next twenty years.
Newell's theory has three main themes:
As we show in the companion paper [CP], the human factors and system development communities have valued the first two while the artificial intelligence community has valued the third. Most of the key points identified there will be explained very briefly here.
We start with the intuition that knowledge is somehow different from the symbols it is expressed in (kp:int). The notion of levels derives primarily from practice (kp:prac) but it can be supported theoretically (kp:theory). That knowledge and symbols are distinct helps to clarify strategy and contribution of research (kp:resch). Newell claimed that all things that can be said to have knowledge are describable by this theory (kp:scope).
There are a number of distinct levels (kp:dist), not just one or two (kp:pls). Levels are distinct ways in which computers systems might be described (kp:descr) and each level's description is valid (kp:valid) and complete (kp:compl). Levels do not describe different parts of a system, each level describes the same whole system (kp:same), though in different ways. Newell identified five levels and, because of their close correspondence with the levels of linguistics (we found the levels apply not only to computer systems but also to spoken and written media), we argued for the possibility of adding a sixth:
At each level we can talk about a system (kp:sys), a medium (kp:med), components that process the medium (kp:compn), laws of composition (kp:lawc) and laws of behaviour (kp:lawb). This gives us the concepts and vocabularies with which to design and evaluate systems in terms of each level (kp:des) and, because of the laws of behaviour, we can use descriptions at any level to predict behaviour (kp:pred).
Merely aggregating components at a level does not move us up to the next level (kp:agg). Rather, the describer must 'specialize' lower level systems, medium and components by distinguishing what has meaning at the higher level and 'adding' that meaning (kp:meaning). Meaning is important to levels, and Newell discussed how levels are similar to Dennett's  stances (kp:stance), with the knowledge level equivalent to the intentional stance. But, while Dennett thought stances are in 'the eye of the beholder', Newell made a strong ontological claim for levels (kp:ont): "Computer system levels really exist, as much as anything exists. They are not just a point of view." So we cannot just think up intervening levels (kp:nointv).
Thus the distinctness of levels from each other takes the form of an ontological irreducibility (kp:irred). There are several consequences of this. Some things described at one level become invisible at next level (kp:invis). What seems an error at a level is usually explainable and valid functioning in lower levels (kp:err). Another consequence is that the the relationship between levels is many-to-many (kp:m-n) (e.g. the symbol 'a' can be implemented by bit patterns '
01100001' (ASCII) or '
10000001' (EBCDIC), while '01100001' can implement the both 'a' and the number 97), which gives us a number of freedoms (kp:freedom): freedom to choose how to implement at the lower levels, freedom to work at each level without full knowledge of lower levels (kp:detail), freedom in design and prediction. Especially, this means we can draw up knowledge level specifications of (bespoke) systems before knowing how it will be implemented and, conversely, build generic systems without knowing the details of the application. Each level has its own role, its own importance, in design and evaluation (kp:imptnt).
Though distinct, the levels are related, so there should be no 'unbridgeable gap' (kp:rel). The levels form a sequence of dependency: if we have a description at one level then it will always be possible, in principle, to compile a description at lower levels (kp:dep). Each level determines a distinct technology (kp:tgy). So higher level descriptions can always (in principle) be implemented in lower levels (kp:impln). However, there might not be a description at higher levels, so there might be a top level at which a system can be described (kp:top).
The knowledge level was of particular interest to Newell; its system he called an agent and its medium is knowledge. What does it mean to say "the program knows K" [KL:116]? The first part of the answer is the link between symbol and knowledge, which Newell called 'aboutness' (kp:about). Aboutness involves interpretation of symbols (kp:interp), and so there must be an interpreter, which Newell called the observing agent (kp:oo).
Knowledge is about the environment, that is the world outside the system (kp:env). So the knowledge level is intimately bound up with the application of the computer system (kp:appln), and can move us away from techno-centric positions. Variety at the knowledge level comes from content (the diversity of the environment), rather than structure as it does at lower levels (kp:variety). Because of this, and because knowledge is generative (see below) "Knowledge of the world cannot be captured in a finite structure" [KL:107] and is potentially infinite in scope (kp:inf). So, unlike symbol level programming, knowledge engineering is no routine process, and knowledge engineering methodologies must be iterative rather than linear  (kp:kgacq).
But if we obtain a piece of knowledge from an agent, we do not know whether the agent has a symbol for this in their symbol structure or has been inferred therefrom. So the second part of the answer is that knowledge is generative. Newell saw knowledge as a competence (kp:comp) which is more than is represented in their symbols (kp:genv). For this to be so, knowledge is active functioning (kp:func) rather than passive state or commodity. The generative mechanism Newell proposed was deductive logic, and he postulated that the complete answer to what the agent knows is the logical closure of all that is represented inside (kp:clos). So structure and process merge at the knowledge level (kp:proc) (so, as Anderson  points out, the procedural-declarative debate is irrelevant at the knowledge level).
While at lower levels one can describe the behaviour of a system in terms of components responding to local inputs, at the knowledge level behaviour is determined by a global principle (kp:glob). "Knowledge is intimately linked with rationality" [KL:100] (kp:rat), so Newell proposed that the law of behaviour at the knowledge level is the principle of rationality (kp:por):
"Actions are selected to attain the agent's goals" [KL:98]
on the basis of what the agent knows. Notice that goals, not deductive inference, are the driving force of this rationality (kp:goal). Newell argued that the proper role of logic is not as a symbol level representation language, but as a knowledge level means for analysing the world (kp:logic). The principle of rationality links together the three knowledge level components, goals, actions and bodies of knowledge, but Newell recognised that it needs extending (kp:extn).
While behaviour is determined when described at lower levels (kp:lwrdet), at the knowledge level it is not determined (kp:nondet) and cannot be predicted precisely (kp:nonpred). Newell tried to explain how determinacy is lost between the symbol level and knowledge level by saying that the knowledge level description usually has to be augmented with some symbol level description if behaviour is to be predicted, so the knowledge level is 'radically incomplete' as a level of description (kp:ri).
Over the years, the notion of levels has been misunderstood and sometimes wrongly used. Newell remarked "some people have told me that 'The knowledge level' is rather hard to understand" [RKL:33]. In some places Newell's own treatment in [KL], and several of his concepts, remain unclear. This is partly because his paper takes the reader on a journey rather than presenting crisp theory and examples. There are a few inconsistencies (for example, his different descriptions of the symbol level in tables 1 and 2). He does not give full examples of each of the levels, so it is easy to mistake what is included in each. The two themes of aboutness and behaviour are interwoven, and it is hard to see how they relate. Radical incompleteness is a confusing concept. Many such confusions could be removed by a clearer exposition of his levels - which we hope to have achieved in the 'rational reconstruction' of Newell's theory that we have attempted in the companion paper [CP] and summarized above.
Normally knowledge is about something outside the computer system (kp:env). But confusion can arise (especially seen among my students) when knowledge is of symbols or lower levels. For example, the Auscor corrosion predictor system  'knows' that acidity can cause corrosion. But it can also be said to 'know' that, on a dialog box, the 'OK' and 'Cancel' buttons go at the bottom. Not only would Newell's theory not differentiate between these two, but, with its emphasis on visible behaviour, it might focus the observer's attention on the trivial knowledge at the expense of the important. We need grounds for differentiating such types of knowledge (and this requires philosophy). A similar confusion can arise when the application purpose of the computer system is to process symbols, e.g. an editor. It also relates to the special case of self-knowledge, cited in [CP], when the 'environment', of which the agent has knowledge, is itself, either its own symbol structure (e.g. adaptive or self-reflective systems), its own bit level resources, etc.
The confusion of the themes of aboutness and rational behaviour in Newell's paper has led to misunderstandings in debates, such as that between van Heijst et. al.  and Guarino . The former focused on aboutness while the latter focused on rationality. The danger, if we do not resolve this confusion, is that two (or more) camps might emerge that having different understandings of the theory, and valuing different things in it, will use incompatible concepts even when using the same terminology. The theory will then split into fragments, debates will become unhelpful and research will be hindered. We must not only clarify the difference between the two themes, as we have done here, but ask what is the relationship between them. Which needs which, if any? To answer this in a principled manner requires philosophical treatment, both to understand what aboutness and the relationship between other levels are (see below), and to understand how this relates to behaviour.
Our difficulty is compounded by Newell's later introduction [UTC] of a third theme of adaptation, which distinguishes 'representational systems' from symbol systems, on the one hand, that provide distal access, from knowledge systems on the other, whose behaviour is determined solely by content of the representations.
Since Newell conceived of the knowledge level much has happened, not only in artificial intelligence but in information technology and systems as a whole. Technologies have become important that were non-existent or unimportant at the time - computer games, multimedia, virtual reality, computer music and art, desktop publishing, the Internet, case based reasoning, neural nets - and even WIMP user interfaces. It would be useful to explore how his notion of levels can be applied to new technologies.
In the main, it can be applied surprisingly well if we focus on the two themes of irreducibility and aboutness. Just as levels can be applied to multimedia, as discussed in [CP], so they can be applied to web sites, games, computer music and art, etc., all of which involve a close relationship with human thinking, communication and interpretation. The way we understand each level might need widening (e.g. bit level includes pixels and gestures) but few other changes are needed.
But some technologies have emerged in which Newell's contention (kp:dep) that all systems describable at the knowledge level are also describable at the symbol level breaks down: the symbol level seems to be missing. Neural nets work subsymbolically and yet provide results that can be meaningful at the knowledge level. Likewise when a virtual reality system creates a mountain landscape it can be recognised at the knowledge level, as a landscape, but there might be no symbols in the whole screen. Thirdly, the concept of direct manipulation in user interfaces is being redefined to map the lexical (bit) level directly to the semantic (knowledge) level, bypassing the syntactic (symbol) level [49, 6]. Newell's theory must be refined either to account for these technologies or to recognise why it cannot.
Finally, the advent of the Internet poses a problem. When a computer is connected to the Internet, does its knowledge level include or exclude all the knowledge on the Internet? What of the logical closure of it all? What about dynamic connections? Such dilemmas occur, in fact, at all levels, though they are most acute at the knowledge level. We must find a principled way of defining what we make our level descriptions of; this is likely to require a philosophical stance that addresses purpose.
Not only has technology advanced, but two new perspectives on information technology have come to importance. Common to both is the increased theoretical importance attached to human beings involved with the computer system.
One of these is human factors, which sees the human being, not the automaton, as of prime importance. Kammersgaard  discusses four perspectives. The system and dialog-partner perspectives see the computer as an agent of essentially equal status with the human being, the difference between them being on the individual-group axis. The tool and media perspectives (again differentiated by the individual-group axis) see the computer as an artifact employed in human functioning, with emphasis on human tasks  and benefits [41, 4] or as a medium through which humans communicate. Artificial intelligence and early human computer interaction such as  work from the earlier perspectives, and Newell's theory was developed from a systems perspective. But the knowledge level is arguably more compatible with the tool and media perspectives, especially because of kp:oo, kp:interp. The system and dialog-partner perspectives require the behaviour theme while the tool and media perspectives require that of aboutness. It seems that Newell's theory wants to span all the perspectives, but it might need some refinement in order to do so comfortably.
The other shift in perspective has been from positivism, that emphasises rationality and deterministic relationships, to interpretivism, that emphasises human interpretation, as espoused by [16, 43, 58], and more recently to criticalism, that emphasises power relations and emancipation . Newell's paper, though born of a positivist outlook, possesses elements that might allow it to be compatible with interpretivist perspectives - especially interpretation (kp:interp), the observing agent (kp:oo) and the non-determinacy of the knowledge level (kp:nondet). Newell tried hard to explain them in positivist terms but the result is less than convincing, as we see below. We need to ensure Newell's theory is at least compatible with the new perspectives if it is to make useful contributions for another 20 years. But the paradigm of the detached observer, that echoes through Newell's approach, is being replaced by that of the involved participant, so we might have to modify our view of how the describer determines the knowledge of the observed agent, and thus maybe rethink radical incompleteness. We must understand the perspectives which, in turn, requires philosophy.
That levels are a "fact of nature" and not "in the eye of the beholder", as Dennett believed, was important to Newell; he made this ontological claim (kp:ont) at least three times in the original paper and reiterated it ten years later.
While it might be easy to make such a claim from within a postivist perspective, from a more interpetivist perspective, ontological claims are suspect because they smack too much of essentialism, which is out of fashion in many circles today. So we must either reject his claim or justify it. What we cannot do is ignore it - even though almost everyone who has referred to the knowledge level thus far has done just that [RKL:33].
Unfortunately, Newell did not justify his claim. The companion paper [CP] discussed several reasons why it is important, such as that it is the basis for the very notion of levels as irreducible and yet related. It validates our intuition that knowledge should be studied as knowledge rather than as something that emerges from symbols, psychology, logic or sociology, and thus stimulates the whole enterprise of knowledge engineering. But these do not justify the claim; we must do so from philosophy.
The ontological claim, however, is not as simple as it seems. To say that the levels exist is to say something about the nature of levelness. The claim must include a unifying principle by which levels exist and the conditions that pertain across all levels and the differences between them are accounted for. It must also give us criteria by which we may judge new candidate levels, such as our tacit level, Jennings'  social level and Newell's own Problem-space Computational Model [UTC] - and, where there is a conflict with the current suite, to decide on principled grounds whether it is the candidate or the suite itself that is wrong.
Brooks  contends that knowledge level theories "too often chase fictions". Domingo and Sierra  speak of "This gap between knowledge level analysis and the implementation ... the source of the main criticisms to this approach to KBS building." There are, they say, "philosophical suspicions that a relation from such an abstract description and a concrete implementation framework could never be found." The problem lies in the irreducibility of the knowledge level to the lower levels (kp:irred), in that the very freedom this affords leaves the analyst and designer with little guidance on how to implement the desired system. The same could be said of the gap between other levels. If we are to bridge such gaps we need to understand the nature of the relationship between the levels, and how such an understanding can help us formulate practical methodologies and guidelines for implementation. In the end, though, this might be a matter of methodology.
Dietterich  pointed out that Newell's discussion of the knowledge level did not address the possibility of knowledge of a system changing, as happens in learning, knowledge acquisition and forgetting. He sought to remedy this by discussing learning at the knowledge level. While his approach might not be the only valid one, it is carefully argued and many have referred to it. Because of the postulate of logical closure (kp:clos), whose problems are discussed below, Dietterich comes up with the surprising result that some types of learning cannot be adequately described at the knowledge level, so that Dietterich's approach does not accord well with our usual notions of learning. But other approaches are possible. As discussed above, and more fully in [CP], we could treat induction algorithms as knowledge. We could also question Dietterich assumption that learning is goal-directed behaviour. This work has still to be carried out - as has addressing forgetting or negative changes in knowledge - and we will not discuss this further.
Gaines  points out that "Newell's notions of rationality, goals and knowledge have no epistemological content and are circularly derivable from one another as definitions of what it is to adopt an intentional stance." We need to ground this circular definition in something outside itself. Philosophers have produced various theories of knowledge but our specific task here is to find one that is in harmony with Newell's theory. I suggest the key to this is the symbol-knowledge relationship that Newell calls 'aboutness'. But "the knowledge level does not itself explain the notion of aboutness; rather, it assumes it." [KL:123] He discusses one thing that representation can achieve, namely distal access, but this does not give us what we need, an actual account of 'aboutness'. Dietterich  suggests that it could be tackled in terms of possible-worlds semantics  and Mingers  tries to argue that it emerges from Maturana's autopoiesis, but such attempts are rather complex and are inconsistent with Newell's ontological claim (kp:ont). We will seek an account in the philosophy outlined later.
Newell proposal that the law of behaviour at the knowledge level is the principle of rationality (selecting actions according to goals and knowledge) has a number of problems. Though in line with the logicism and positivism of Newell's time, it is out of tune with today's interpretive perspectives, and also has a number of inherent problems.
Newell claimed that it helps an observer predict behaviour of an agent without knowing their symbol level because the observer can know the environment of the agent and hence know what knowledge it has, can find out its goals (either by asking it or by applying the principle to its past actions), and then can predict future actions from the same principle. This is possible because both goals and knowledge of the environment are relatively stable. But Dietterich  claims the predictions this principle allows us to make are not only weak but extremely sensitive to the knowledge the observer has about the goals and knowledge of the observed agent. Small inaccuracies in the observer's knowledge of the agent, especially of its goals, are likely to be magnified into large errors in prediction. The principle of rationality seems too blunt an instrument for making knowledge level predictions, even though it might be a useful starting point.
Goals themselves present a problem. Goals are not always as clearly defined as Newell's treatment would imply. Sometimes an observer might not know what goals can be attributed to the observed agent. The observed agent itself might not be aware of its own goals and thus not be able to apply the principle of rationality properly. Furthermore, since the purpose of the law of behaviour at the knowledge level is to account for the link between action and knowledge, then we need some way of accounting for what goals an agent has. The principle of rationality cannot do this on its own.
Habermas  speaks of five different types of action, of which only the first, teleological action, is covered by Newell's principle of rationality. The others (strategic, normative, dramaturgical and communicative) speak of a diversity of human behaviour that transcends rationality. So we might wish to devise additional principles to work alongside the principle or rationality. Why not have a Principle of Harmony, for instance: "The action will be selected to harmonize with the agent's existing knowledge." From Winograd and Flores  we might wish to add a Principle of Breakdowns (rendered for example as "Knowledge only becomes salient when the unexpected happens"). From Habermas  we might add a Principle of Self-Reflection to Identify Hidden Presuppositions. The Principle of Justice ("an agent will select actions by which other agents receive what is their due") or the Principle of Self-Giving Love ("an agent will select actions that benefit the other even if sacrifice is involved") would seem to govern at least some human behaviour, especially that which we would applaud (and which we might hope that intelligent agents will emulate). Indeed Newell himself gives precisely an example of self-giving love, in his recitation of the story of The Lady and The Tiger (see below).
Newell did recognise the principle needs to be extended [KL:103], and made a few suggestions. Carley and Newell  extend the principle of rationality several steps to incorporate emotion and cultural-historical knowledge, but it is likely even this is insufficient. Some behaviour does not involve goals at all, but is more of a response to the environment or norms. Biological processes, which include autopoiesis , are of this kind and one might argue that true sacrifice should be an attitude rather than a goal. To Habermas , presuppositions are ever-present and interests underlie all knowledge. So, while the principle of rationality might be retained, it cannot stand alone and must be accompanied by others that are of equal standing and cannot be reduced to it. We could extend it by adding principles that generate goals and ones that bypass them. But how are we to identify these new principles? While reflection or empirical study might provide useful suggestions, a philosophical analysis can help, as we show below.
In line with his emphasis on rationality, Newell postulated logic as the generative process in knowledge, and the logical closure of what is represented as what the agent can be said to know (kp:clos).
"[Logical closure] is good for predicting that a person can find his way to the bedroom of his own house, for predicting that a person who knows arithmetic will be able to add a column of numbers. And so on, through much of what is called common sense knowledge." [KL:110]
But the postulate of logical closure gives rise to several problems. Dietterich , for example, suggests logical closure makes much knowledge unimplementable, because it would be too large (often infinite). Aamodt and Nygård  say that it makes the knowledge level difficult to use for practical purposes. But it also has fundamental problems.
Logical closure is simply counter-intuitive, in that actual behaviour of people does not display such omniscience as the postulate would suggest. Newell admitted this - people who might know P and that P implies O do not necessarily know O  - and explained departure from logical closure by effects from lower levels such as resource limitations. But resource limitations might not be the only reason for failure to know O.
Some factors that we would expect to be able to identify and discuss at the knowledge level are made invisible by the postulate of logical closure:
Is it possible to remove, or replace, it as the explanation of what an agent knows? An indication that Newell might have welcomed this is found in his comment on Dietterich's work, "is restricted ... by deductive closure" [RKL]. To refine Newell's theory appropriately will take work, and the philosophical treatment below might point us in the right direction, but we can make the following preliminary suggestions.
First, it seems useful to distinguish more clearly between k1, what the symbols within the agent are about, k3, the total of what it is possible to know from k1 (the logical closure in the current version of the theory), and k2, the knowledge the agent actually is using in the situation of interest. K2 is, of course, a subset of k3 and a superset of k1. The behaviourist would argue that an observer can never delineate k1 precisely so it should not enter our theories, but Newell dismisses this stance. K1 should be an element of a refined theory for several reasons:
Second, we might take Anderson's  line that knowing how to infer, induce, etc. are themselves knowledge, and visible at the knowledge level, and work these into the theory, not relegating them to the symbol level. How this can make the results of inductive learning visible, even with logical closure, is discussed in [CP]. This relates also to the issue of self-knowledge, discussed above.
Third, we should replace deductive logic as the generative mechanism (in line with extending the principle of rationality). How this might be done, and a basis for doing so, is discussed later. Logical closure might then be retained as a special case, if desired.
Newell's postulate of the 'radical incompleteness' of the knowledge level (kp:ri) is related to that of logical closure and likewise gives us problems. It is tied up with Newell's claim that while behaviour described up to the symbol level is determinate (kp:lwrdet), that described at the knowledge level is not (kp:nondet), and so behaviour at the knowledge level is non-predictable (kp:nonpred).
Newell showed the need to postulate radical incompleteness with Stockton's  Lady and Tiger story. The lover of a princess must choose between two identical doors, one of which leads to a tiger, the other to a beautiful lady. The princess finds out which leads to which, and by surreptitious signal indicates which door her lover should open. However, the princess had found out that the lady door would lead, not to herself but to her chief rival. Which door did she tell her lover to open? Did she send him to death or to life with her rival? "Our knowledge-level model of the princess," said Newell, [KL:104], "even if it were to include her complete knowledge, would not tell us which she chose."
Use of this example strongly suggests that Newell had real non-determinacy in mind, but the actual explanation he gave in [KL] seems to imply that he treated non-determinacy as a mere artefact of our lack of knowledge of the agent. Her actual choice, he implied, is dictated by the precise operation of her symbol level algorithms acting on the symbol structures etc. that represent her knowledge. If we knew what these were, we could predict her choice precisely. But because the actual algorithm is invisible at the knowledge level, no purely knowledge level description can predict which she chose, so we must augment the knowledge level description with some symbol level description. Further, we cannot even infer her symbol structures or algorithms from past behaviour because of the generative nature of knowledge (kp:genv, kp:clos). A knowledge level description is, therefore, 'radically incomplete' (kp:ri). But that explanation has flaws.
The program finds the value inexplicably changed, seemingly randomly (SL). The symbol level description, which includes only steps 2 and 5, is radically incomplete.
Newell seemed to see all or nothing when it came to prediction: either total predictability or radical inability to make any prediction at all. But if we relax our demands of prediction, so that it provides explanations and expectations rather than absolute predictions, we might be able to move towards a solution. We can use what Newell called common sense knowledge. We can never tell which the princess will choose, but if we know the princess is self-giving and wanting the best for others, we would expect her to choose the lady door whereas if we know she is proud or jealous we would expect her to choose the tiger door. These knowledge level predictions are useful and, arguably, of higher quality and usefulness than any afforded by resorting to symbol level augmentation. But, on what basis is such a move valid? To answer this requires, yet again, philosophy.
Radical incompleteness means that, unlike lower levels, the knowledge level cannot be relied upon to fulfil its whole duty as a level of description, that is to furnish a useful description of the system that is complete at the level (kp:compl). This creates a dissonance in Newell's theory of levels. In fact, Newell identified four 'surprises', ways in which he claimed the knowledge level to be different from the others. Some difference is to be expected because each level is distinct, but the implication is that it is far greater than the differences that exist between lower levels, and that we must make special exceptions in the underlying principle of levelness for the knowledge level.
It would be unfortunate if this is so because it disrupts the theory of levels and makes it inelegant. Having to remember what these exceptions are makes the theory of levels more cumbersome to use. There is also a danger that the knowledge level becomes separated off from the others and assumes almost magical qualities - something that Newell specifically wished to avoid [KL:90].
We might eventually be forced to accept that the knowledge level is an exception in the suite of levels, but we should not give up our search for harmony too easily. In fact, when we examine Newell's 'surprises' more closely - and especially with help from the philosophical underpinning below - the disruption is reduced.
Surprise 1. Variety of behaviour comes from content at the knowledge level, but from structure (how components are assembled) at lower levels.
Both parts of this are questionable. At lower levels, content (e.g. values of symbols, bit patterns in registers) exhibits some variety. At the knowledge level structure within the knowledge itself is important. Examples might include Clancey's  heuristic classification, or links to cultural assumptions, as illustrated in Newell's own example [KL:98], "She knows where this restaurant is and said she'd meet me here; I don't know why she hasn't arrived." So variety at all levels comes from both structure and content. But there is one difference: structure at the knowledge level is within content because of the central part played by 'aboutness'. So this 'surprise' comes from the very nature of the level itself, and is no longer a surprise.
Surprise 2. Behaviour at the knowledge level is determined by global principle, but at lower levels, by local processing of the medium by each component.
Later we will see it is possible to argue that behaviour at all levels is determined by global principle, and apparent local processing can be explained in such terms.
Surprise 3. Behaviour at the knowledge level is not determinate, while at lower levels it is. Lower level descriptions are complete while, often, a knowledge level description must be augmented by symbol level description.
As discussed above, this is a serious disrupter. But we discuss later how non-determinacy can be accounted for without the postulate of radical incompleteness, nor that of logical closure, and the knowledge level resume its full duty as a level, so the problem disappears.
Surprise 4. Knowledge is active while the medium at lower levels (voltages, bits, symbols, etc.) is passive states changed only by components.
As with the first surprise, we must question both parts. A list of theorems written on a blackboard, Newell argued [KL:105], is a symbol structure which could equally well reside in computer memory cells (BL) and voltages (CL). These all have physical structure, but "all this fails at the knowledge level. The knowledge cannot so easily be seen, only imagined as the result of interpretive processes." But can it not? On one hand, we can indeed see knowledge: I walk into the room and see five theorems (KL), not a list (SL). On the other, interpretation is involved in seeing what is on the board at all levels. Unfortunately, with "Though the seeing is a tad harder ..." Newell passed over the very crux of the matter. Moreover, lower levels can indeed be active, with dynamic things like animation and sound being just as good as symbol as those written on a blackboard, while some knowledge is static.
We will see later that philosophy provides at least one relatively simple, single harmonizing principle for levelness into which the above 'surprises' can fit, given the modifications just indicated.
Newell's theory would disallow any disembodied ('spiritual') knowledge because he claims that knowledge needs physical bodies to implement it. We do not discuss this here, but to do so would require clarification of what is meant by 'spiritual'.
As we can see, apart from some minor confusions that can be cleared up by better presentation, and the need to undertake research to extend Newell's theory to account for knowledge change, most of the problems require philosophical treatment:
We now seek such a philosphical underpinning.
Newell himself discussed links with philosophy, but only very briefly. After touching on the relevance of the philosophic notions of mind, knowledge, certainty and belief, he discussed the links his ideas have with Brentano's  concept of intentionality and Dennett's  application of this to artificial intelligence, as the intentional and the subpersonal stance. The stances are equivalent to the knowledge level and symbol level, but there are significant differences in detail, and Newell called for closer analysis [KL:123].
While Dennett's and Brentano's ideas can usefully underpin some of Newell's points, they are insufficient to underpin the whole theory. There is much philosophy beyond positivist Brentano. We must seek a broader philosophical underpinning - one that supports every key point and can explain why it is a coherent part of the whole theory. We want to do more than just find quotations or concepts from a selection of philosophers to attach to each key point. Rather, we seek a coherent and comprehensive philosophical framework that can support all the key points in harmony.
But it turns out that it is surprisingly hard to find a coherent underpinning for Newell's theory amongst conventional philosophies.
To understand why this is so, we need a broad picture of philosophy, such as painted by Tarnas . He shows there have been two main streams in philosophy, and culture as a whole, [ibid.:366]:
"From the complex matrix of the Renaissance had issued forth two distinct streams of culture, two temperaments or general approaches to human existence characteristic of the Western mind. One emerged in the Scientific Revolution and Enlightenment and stressed rationality, empirical science, and a skeptical secularism. The other was its polar complement, sharing common roots in the Renaissance and classical Greco-Roman culture (and in the Reformation as well), but tending to express just those aspects of human experience suppressed by the Enlightenment's overriding spirit of rationalism."
His two poles are those of Determinism and Freedom. The former explains things in terms of deterministic laws of Nature and values control, predictability and theory, while the latter explains in terms of the free human ego or spirit and values creativity, existential experience and intuition. They take different forms and flavours in different communities and eras, and Tarnas speaks of scientific rationalism, positivism and Enlightenment thinking clustering around one, and Romanticism, Existentialism and what he calls perspectivism around the other.
The success of science and rationality in overcoming the 'superstition' of the mediaeval era and in helping us understand physical phenomena is well known. Newton, Descartes, Leibniz and the like gave us new routes to knowledge of the world that is reliable and certain, an ability to predict, and the capacity to control. Logical positivism later arose as a major scientific paradigm. Science's "claims to valid knowledge of the world ... continued to seem not only plausible but scarcely questionable" [ibid.:355].
The power came from the scientific method and by reducing complexity into simple parts. This strategy was followed in artificial intelligence until well into the 1980s, and Brachman and Levesque  trace its roots to Leibniz's idea "that an analysis of ideas could be devised, whence in some combinatory way, truths could arise and be estimated as though by numbers." Newell's paper was written in this atmosphere and culture, and we can see several elements of his theory that accord with it, notably his linking knowledge intimately with rationality (kp:rat). His proposal that the law of behaviour at the knowledge level is the principle of rationality (kp:por), and that what an agent knows is a Leibnizian combinatory logical closure (kp:clos) are obviously of the rationalistic stream, as is his claim that at lower levels behaviour of the system arises from the determined behaviour of simple components (kp:lwrdet). In this way scientific rationalism can underpin several of Newell's key points.
But it cannot underpin all. The central importance of interpretation to the knowledge level (kp:interp) and its non-determinacy (kp:nondet) are not only rejected but cannot even be recognised. Positivism presupposes a detached scientific observer who not only has no effect on what is observed but is also not part of the result of observation, but the knowledge level cannot exist without the observing agent (kp:oo).
Philosophies of the other stream, Romanticism, Existentialism, Phenomenology, perspectivism, etc., arose partly in reaction to scientific rationalism. Romanticism "affirmed the inexhaustible drama of human life rather than the calm predictability of static abstractions" [54:367], valuing things that the rationalistic stream ignored such as spirituality, imagination, emotion, will. So whenever "the classical Cartesian-Newtonian cosmology gradually and then dramatically broke down ... the long-established certainties of classical modern science were radically undermined" [ibid.:355] and "fundamental Kantian a-prioris - space, time, substance, causality [which many papers in artificial intelligence assume] - were no longer [found to be] applicable to all phenomena" [ibid.:359], philosophies around the Freedom pole would grow.
Nietzsche, who stood at the end of Romanticism and at the beginning of perspectivism (or postmodernism), said: "Against positivism, which halts at phenomena - 'There are only facts' - I would say: No, facts are precisely what there are not, only interpretations." So Newell's emphasis on interpretation, and active interpretation at that, for the knowledge level (kp:interp,kp:func), is therefore of this stream. So is his claim that behaviour at the knowledge level is fundamentally non-determinate (kp:nondet). While scientism conceived of the world as an atomistic machine (whether material or rational) observed and controlled by the detached human being, the Romantic vision perceived it as a unitary organism. That the knowledge level is about the environment rather than the system itself (kp:env) reflects this. "Rather than the distanced object of sober analysis, nature for the Romantic was that which the human soul strove to enter and unite with in an overcoming of the existential dichotomy, and the revelation he sought was not of mechanical law but of spiritual essence" [ibid.:367]. Though nowhere spiritual, Newell's dyadic integration of observer with observed (kp:oo) is obviously closer to the Freedom pole.
Thus we can see Newell's theory swinging between the two poles as it develops. First, at the lower levels behaviour is determined (kp:lwrdet). When the knowledge level is introduced, with its interpretation (kp:interp) it moves towards the Freedom pole. Tying knowledge to rationality (kp:rat) draws us back to the scientistic view. The non-determinacy of the knowledge level (kp:nondet) moves us towards the Freedom pole again. But Newell tried to pull us back to the Determinism pole by his explanation of it based on radical incompleteness (kp:ri). But the knowledge level's close involvement with the environment (kp:env,kp:oo) moves us towards the Freedom pole again, for which "nature was ... a translucent source of mystery ..." [ibid.:367].
Newell's ontological claim of a suite of five (or more) levels is problematic for both poles. At first it would seem to be supported by the Determinism pole because philosophies around this pole tend towards philosophical realism, but, as we shall see below, it cannot support Newell's pluralism (kp:pls). Freedom pole philosophies can support pluralism but are antithetical to any ontological claim (kp:ont) because, based on the freedom of the human ego, they tend to adhere to philosophical nominalism.
The question is, then, what should we do? We could abandon the search for philosophical underpinnings altogether, as Gaines  does. But we listed earlier many problems that require a philosophical treatment (and the way Gaines subsequently develops his ideas exhibits serious flaws that a philosophical underpinning might prevent). We could try to push Newell's theory to one pole or the other and hope that somehow, after sufficient debate, we can find a way in which all the key points can be underpinned. But there is no guarantee that this will ever be possible and any attempt is likely to be extremely complex, even arcane. We could develop two versions of Newell's theory, one that omits all traces of Freedom and one that omits all traces of Determinism. As discussed above, something like this has happened to a degree, and it led to confused debates in which propopents argue at cross purposes. No! We do not want to give up our search for a coherent philosophical underpinning too easily.
The root of the problem - and the solution - is deeper than we might have expected; see Fig. 1. The theory of the knowledge level illustrates what Tarnas [54:375] calls "a complex bifurcation of the Western outlook". As we have already mentioned Newell wrote his paper in a rationalist, postivistic culture that admits of no human interpretation. That is, he wished to:
But he was not at liberty, when theoretically addressing the link between the symbol and knowledge levels, to postulate a genuine free act of interpretation and had to seek some kind of link that could accommodate determinism.
Trying to embrace both poles, trying to account for something of one pole from the perspective offered by the other, leads to an antinomy - a fundamental inconsistency that cannot be resolved by means of theoretical analysis. Though several thinkers have attempted syntheses of Determinism and Freedom (Tarnas discusses Goethe, Hegel and Jung), it is not clear that any of these attempts succeeded. On Hegel's attempt, Dooyeweerd [23:84-5] commented,
"Against Hegel's synthetical dialectic which attempted to think together the antithetic motives of nature and freedom, Proudhon directs the verdict, earlier pronounced by Kant and later repeated by Kierkegaard: 'L'antinomie ne se reisout pas' (The antinomy cannot be solved)."
Trying to embrace both Freedom and Determinism in a theory, as Newell's theory demands, cannot be done. This is because our presupposition that they are mutually exclusive is too strong and lies deeper than theoretical thought can reach. To overcome an antinomy of this kind requires that we first move to a different philosophical starting point that adopts different presuppositions.
After several years of using and valuing Newell's theory and in particular its practical usefulness and its intuitiveness, this author came across a relatively new stream of philosophy that might serve our purposes. This was the philosophical stream proposed by Dooyeweerd  and it questions the presuppositions. He showed that it was because of those very presuppositions that antinomies arise. He also showed that they arise only in theoretic thought, and need not trouble the practical situation of everyday living. However, to understand these, we have to go back 2,500 years, to the very foundation of Western thinking itself.
"All else is a mere footnote to Plato," remarked Whitehead , pointing to the way most Western thinking over the last 2,500 years has largely rehearsed the issues that Plato brought to our attention, albeit in different guises and contexts. For example, the shift in perspective in information systems, from positivist to interpretivist, is just one variation on the Determinism-Freedom theme. The footnote is presuppositions: presuppositions that Plato, and other thinkers of his time, made and that we still make today. We will show, by summarizing parts of Dooyeweerd's  arguments very briefly, that it is these presuppositions that have prevented our underpinning Newell's theory (as well as generating other problems that do not affect us).
One of the presuppositions is made quite explicit by Gaines [25, his italics]: "The most fundamental properties which we impute to any system are its existence and persistence over time." Existence can be imputed to not only a physical object but also a concept, a process, a symphony, a government - and a level. But Heidegger  challenges our lack of perplexity over the meaning of Existence (Being), and Hirst  gives us reason to be 'perplexed' in his discussion of existence assumptions. While the existence of physical objects might seem unproblematic, the 'existence' of other types of thing seems to raise questions. Does a living organism 'exist' after it dies? Concepts can be said to 'exist' in the thinker's mind, but what happens after that person dies, or even forgets the concept? Suppose the concept is committed to paper before the person dies but nobody ever reads the paper; does the concept still 'exist'? What about Nielsen's Fifth Symphony? Does justice 'exist'? The patch of sun on the sea? The accident that was prevented by somebody's prompt action yesterday: does that 'exist'? Do fictional figures like Gandalf 'exist'? Finally, what about impossible things like a square circle or the present King of France; could they 'exist'? Clearly the notion of Existence is not as straightforward as we assumed. In particular, what does it mean to claim, as Newell did [KL:99], "Computer system levels really exist, as much as anything exists"?
Behaviour and functioning seem to derive from Existence if we also presuppose that the focus of theoretical analysis should be on Entity rather than Law. This presupposition is implicit in Gaines' statement, and he quotes Peirce  ("the first germ of law was an entity") to argue that laws of behaviour emerge from Entity. That is, tendency to behave in certain ways is to be found within the entity itself, or, rather, on those of which it is composed. No wonder Newell found it a 'surprise' that knowledge level behaviour comes from a global principle (kp:glob).
These two presuppositions are rooted in a deeper one from Plato's time, that there is something that 'exists' in our temporal cosmos or experience on which everything else depends and which itself depends on nothing else . Dooyeweerd  called this Immanence Thinking. Most of us hold a similar presupposition today, the main differences between us, that we have been debating since Plato, are
To (a) two main types of answer have been given since the Renaissance: either something determinate and outside humanity or the free human ego, spirit or interpretation - the Determinism and Freedom poles we met above. To (b) the philosophical realist would answer "Yes, we can observe and reason about it", the philosophical nominalist would answer "No, we create our own reality" and the so-called critical realist would answer "Not really, but it is there". But whatever answers we give to these questions, this basic presupposition leads to problems.
One is reductionism, our reducing all to one self-dependent thing. Whatever we hold to be self-dependent (be it Reason, Economics, Evolution, Interpretation, etc.) becomes the grounding principle from which all else is derived or emerges, the basis on which all else is explained and, ultimately, the most important thing, which must be protected and to which all else might be sacrificed. Dooyeweerd argued that such reductions are ultimately religious, rather than theoretical. Under reductionism, coherence and diversity become divorced, so Newell's ontological claim for a plurality of levels that are distinct and yet related (kp:ont, k;:irred, kp:rel) must be denied. There are two types of reduction. Either we reduce all levels to one (e.g. the physical), making them mere abstractions of it, or we reduce all to human interpretation and levels become arbitrary stances taken by an observer. As a result, either we deny true diversity of levels or their coherence dissolves into fragmentation.
Another problem is that presupposing the primacy of Existence gives no way by which true Meaning can be accounted for. (We can see echoes of this in controversies over consciousness.) Though many thinkers have tried, via linguistics or Existentialism for example, no way has been found by which Meaning can derive from Existence. But Meaning seems to be central to Newell's levels, in that each is defined by a certain type of meaning (kp:meaning).
A third problem, Dooyeweerd  argued, is that Immanence Thinking leads inescapably to dualistic sub-theoretical ground motives that force thinkers to set two ideas in polar opposition. The dialectical process by which this happens is not explained here (see ) but Dooyeweerd demonstrated that this was so in a myriad of thinkers from before Socrates, through Plato and Aristotle, through Augustine and Aquinas, through the many thinkers of the Enlightenment, examining Kant in particular depth, and well into the twentieth century. He showed that, over this period, three major dualistic ground motives have held sway: Form-Matter from Plato's time to around 500 AD, Nature-Grace until 1500 AD and Nature-Freedom (or Determinism-Freedom) since that time. Ground motives are deeper than world views: their poles generate world views. He also showed that antinomies are inherent in all Immanence Thinking. So we, trapped within the Nature-Freedom ground motive, cannot bring together the polar opposites even if we attempt or wish to.
What this means is that if we wish to find a truly coherent philosophy that can support all of Newell's key points,
we must escape the Immanence presuppositions. (Might we return to an earlier ground motive? Those lead to other problems, driven by different polar oppositions, e.g. the Form-Matter opposition would not allow us to link the physical with the abstract.) Dooyeweerd argued that the problem is not which (dualistic) ground motive we adopt but rather with the presuppositions of Immanence Thinking itself.
Not wanting merely to demolish, Dooyeweerd accepted the challenge of constructing an alternative framework that escapes the presuppositions of Immanence Thinking. He started from a different type of ground motive - the Hebraic one of Creation-Fall-Redemption. This allowed him to presuppose Meaning rather than Existence as the fundamental property of all around us, and Law rather than Entity as the focus for theoretical analysis, and he showed how a number of the deep problems might be solved. See Table 1.
|Nature of Temporal Reality
|Dependent on Other
|Focus of theoretical analysis
Instead of Meaning emerging from what we are or do, Meaning is the very ground of what we are and do, and these emerge from Meaning. Instead of a thing having or generating meaning, it is meaning:
"Meaning is the being of all that has been created and the nature even of our selfhood." [23:6, italics in original]
If this is so, Dooyeweerd had to account for distinct yet related types of Meaning and how Meaning can be the ground for existence and functioning, for what Heidegger meant by Being and Time. He proposed a comprehensive pluralistic ontology of Meaning that provides a foundation for coherence and diversity by removing the need for reductionism. We cannot present the whole of Dooyeweerd's wide-ranging thought here, even in outline, but will focus on those parts that provide a useful underpinning (kp:theory) for Newell's concept of levels, and the knowledge level in particular. We will interleave explanation of portions of Dooyeweerd's thought with their application to Newell's theory.
Just as Newell claimed that his notion of levels had grown out of practice (kp:prac), so Dooyeweerd started, in the Prolegomena of his major work [23:3], not from theoretical analysis, but from everyday experience:
"If I consider reality, as it is given in the naïve pre-theoretical experience, and then confront it with a theoretical analysis, through which reality appears to split up into various modal aspects ..."
('confront' recognises the part that theory and analysis might play with respect to everyday or real life experience; they he returns to the theme of coherence ...)
"... then the first thing that strikes me, is the original indissoluble interrelation among these aspects which are for the first time explicitly distinguished in the theoretical attitude of mind. ..."
So the role of theory is to distinguish aspects . Then he immediately turns to what is perhaps the most visible part of his thinking, his proposal of modal aspects of our experience that account for its coherent diversity:
"... A[n] indissoluble inner coherence binds the numerical to the spatial aspect, the latter to the aspect of mathematical movement, the aspect of movement to that of physical energy, which itself is the necessary basis of the aspect of organic life. The aspect of organic life has an inner connection with that of psychical feeling, the latter refers in its logical anticipation (the feeling of logical correctness or incorrectness) to the analytical-logical aspect. This in turn is connected with the historical, the linguistic, the aspect of social intercourse, the economic, the aesthetic, the jural, the moral aspects and that of faith. In this inter-modal cosmic coherence no sincle aspect stands by itself; every-one refers within and beyond itself."
The aspects are in fact a spectrum of Meaning, differentiated by Time like light is differentiated by a prism. Each has a distinct kernel meaning of its own; see Table 2 but do not attempt to remember all fifteen of them at present! While those who have developed his thinking might modify his suite of aspects [e.g. 19, 31] the notion of aspectuality itself is a major and original contribution, whose characteristics can underpin much of Newell's thought.
|energy and mass
|sense, feeling and emotion
|distinction, clarity and logic
|history, culture, creativity, achievement and technology
|symbolic meaning and communication
|social interaction, relationships and institutions
|frugality, skilled use of limited resources
|harmony, surprise and fun
|'what is due', rights, responsibilities
|self-giving love, generosity
|faith, commitment and vision.
Of particular interest to us is the lingual aspect, because it is the aspect involved in both description and representation of meaning (and therefore knowledge) by symbols. Its kernel meaning is what Newell called 'aboutness' - what a symbol is about - (kp:about), but it goes beyond that to all of linguistics. The symbols can be of any form, in any medium (including those employed in early societies and everyday living). The represented meaning must not be confused with cosmic Meaning, though the two are not unrelated. While Newell held the purpose of representation to be "providing access to a body of knowledge" [KL:114] Dooyeweerd would have seen it as making meaning portable so that it can be detached from its source. The two are not incompatible if we see knowledge as a type of meaning and see that they consider the recipient and originator respectively. While logic can form the basis of a representation language, a survey of the aspects suggests their role to be primarily analytical, in agreement with Newell (kp:logic).
The kernel meaning of each aspect is complex, opening out into many issues and concepts and defining a whole scientific area whose role it is to research those issues (kp:resch). Each such area demands different research methods and issues, and makes a distinct contribution, even offering a basis for interdisciplinary research. Thus the physical aspect involves quantum issues, atomic issues, molecular issues, materials issues, planetary issues, etc. and is researched by physics, chemistry, materials science, etc. The lingual aspect involves most of linguistics, semiotics, etc. What Sloman  calls an architecture relates to the collection of issues of an aspect.
Dooyeweerd's aspects designate ways in which things can be (exist), can function and can be described. We discuss them in reverse order and show how this can provide underpinning to many of Newell's key points. Since all three arise from Meaning, these are integrated; integrating description and behaviour is of especial interest to us.
Each human culture has distinguished concepts and issues in aspects it finds important and developed vocabularies and idioms for them, the more important, the richer they are (with some being multi-aspectual). So each aspect can furnish us with a distinct way of describing an entity or situation. For example, I can be described biotically as a body with life functions, psycho-sensitively as an information processor, analytically as a reasoner, formatively as a goal-achiever, juridically in terms of rights and responsibilities, ethically as someone who needs to love and be loved, and so on, giving up to fifteen possible types of description of the entity. Each such description can make complete sense without any need to reference others.
This, of course, is strikingly similar to Newell's levels, which provide ways of describing (computer) systems (kp:descr) that are complete descriptions of the one system (kp:compl, kp:same). An aspectual description is valid (kp:valid) if the thing described is involved in functioning in that aspect (see below).
The making of a description (as well as the representation of anything) is lingual functioning (as mentioned above). We will see below what this involves. But at this stage it is important to bear in mind that, to Dooyeweerd, no aspect is absolute. All aspectual functioning is fundamentally limited (even though vast), and so no description or representation can capture the full meaning of anything. This is one reason why we might agree with Newell that [KL:107] "Knowledge of world cannot be captured in a finite structure" (kp:inf). It follows that trying to make a description or representation is not a routine process (kp:kgacq).
In fact, we propose here that Newell's levels correspond with Dooyeweerd's aspects (kp:lvls). The distinctness of the aspects accounts for the distinctness of the levels (kp:dist), and their plurality for that of levels (kp:pls). If we compare the kernels of various aspects with what Newell says about his levels the resemblance between them is marked. The correspondence is shown in Fig. 2.
Newell's materials level corresponds almost completely with Dooyeweerd's physical aspect since both are concerned with materials, electrons, fields, etc. The component level, whose concern is components formed out of such materials and assemblages of such to make complete systems, is very similar to Dooyeweerd's biotic aspect (if we can extend 'life functions' to silicon for the time being). The bit level, concerned with Shannonian information, with signals, bits, etc. and their aggregation, corresponds closely with Dooyeweerd's psycho-sensitive aspect. The symbol level, concerned with symbols, corresponds reasonably closely with Dooyeweerd's analytic aspect, whose concern is to identify and conceptualize what is important in a situation; Dooyeweerd sometimes called this the logical aspect. The knowledge level is a little more complicated and spans two aspects, because of its two issues, 'aboutness' and rational, goal-directed behaviour. 'Aboutness', the attachment of meaning to symbols, is of the lingual aspect. Goal-achievement (kp:goal) is of the formative aspect. If we wish to add the tacit level, concerned with cultural or tacit meaning, then this corresponds with the lingual aspect but it anticipates the social aspect, as we discuss below; for now we ignore the tacit level.
|Allen Newell's Level
|Knowledge Level (principle of rationality)
|Knowledge Level ('aboutness')
Fig. 2. Correspondence of Levels with Aspects (Version 1)
The correspondence between levels and aspects is not perfect, of course. To say that the component level is the biotic aspect is something of a liberty since Dooyeweerd was very clear that this is concerned with life functions like respiration, reproduction and the autopoietic maintenance of equilibrium, and we are assuming that the component level processes like manufacture, repair and stability of operation are equivalent; but this has not been properly explored and debated. The bit level seems to have a strong element of resource, yet resources to Dooyeweerd are within the economic aspect. It could also be argued that the digital nature of the bit level should correspond with the distinction-making of the analytical aspect. It might be argued that because the symbol level involves the formation of structures it should correspond with the formative aspect, and it seems unfortunate that while most levels correspond to one aspect, the knowledge level seems to need two.
Nevertheless, for proposals made by two thinkers from completely different cultures, with completely different interests and starting points, the correspondence is strong enough to interest us. Indeed, we will resolve some of these difficulties later (see Fig. 6), noting for now that the knowledge level split corresponds with its two themes. We will find that most of Newell's key points are supported, as we discover more about how Dooyeweerd worked out his theory of aspects.
Another striking similarity is that Dooyeweerd made an ontological claim for the aspects, just as Newell made for levels (kp:ont). But the claim was more precise and philosophically refined, and might help us fill out Newell's claim in ways we discussed earlier. In fact, the aspects do not exist; rather, they pertain. They themselves provide the framework by which all existence is possible (as we see later, in Dooyeweerd's Theory of Entity). They do not depend on us; we depend on them.
Likewise, Newell's emphasis was that levels do not depend on us: "Computer system levels really exist, as much as anything exists. They are not just a point of view." [KL:99] So, if we equate the levels with aspects, we can read this as "Computer system levels pertain, thereby computer systems can exist." In fact, Newell's levels provide the framework by which all computing is possible.
However, strictly, Dooyeweerd's ontological claim was not for his own suite of aspects, which he believed to be in need of refinement, but rather for aspects as such. This is because the process by which aspects are differentiated to us (functioning in the analytic aspect itself) is not absolute, as we discuss later. In the same way, Newell's suite of levels should be open to refinement. However, there are good reasons for thinking Dooyeweerd's aspects are soundly based, and so we can have some confidence in Newell's suite, even if we might refine our understanding of the individual levels themselves.
As mentioned earlier, an ontological claim needs a harmonising principle. Dooyeweerd supplies this: aspects are a spectrum of Meaning, each contributing a different type. Not only does this uphold kp:meaning, but it will expand to provide all else we need, as we consider the coherence and diversity of the aspects, and then functioning.
Dooyeweerd claimed that the aspects are irreducible to each other yet related - just like levels (kp:irred, kp:rel). He discussed two types of relationship: dependency and analogy. As we have argued earlier, in the light of an ontological claim, irreducibility ensures diversity while relationship ensures coherence, giving us a pluralistic suite in which there is no need for reductionism nor fragmentation.
That the aspects are fundamentally irreducible to each other means no aspect can be derived from the others, not even by mutual interaction. We can see immediately that this underpins the irreducible distinctness of the levels (kp:irred). We cannot reach another level by mere abstraction, which Clouser  sees as functioning in the analytic aspect of distinction, but only by the describer bringing in the meaning of the new level (kp:meaning). Things meaningful in one aspect might not be in others and thus either invisible (kp:invis) or an inexplicable error (kp:err).
The dependency relationship means that each aspect depends on earlier ones, so that if an entity has meaning in one aspect it will have at least some meaning in all earlier ones. For example, there can be no physics without movement, space and quantity. Dependency imposes a definite sequence on the aspects. Inter-aspect dependency is part of the unifying principle of levelness that explains why Newell's levels are the ones they are (kp:lvls,kp:ont), and why there are no intervening levels (kp:nointv). The aspect of representation, the lingual, depends on functioning in the earlier aspects:
and each of these also depends on ones before it, as shown in Fig. 3. Newell's levels, tharefore, are precisely those necessary for (knowledge) representation, each lower level being necessary to higher levels (kp:dep).
While, to a given aspect, earlier ones are necessary, later ones are not, so an entity might have a 'top' or qualifying aspect in which it is meaningful. Again, this is very similar to a claim Newell made for levels, that there might be a top level at which something might be usefully described (kp:top). (Note, however, that this dependence is not that of supervenience, which has an additional element of determinism.)
Dooyeweerd's second inter-aspect relationship, analogy, is perhaps unique among ontological suites. In each aspect there are 'echoes' of all the others. For example, causality is inherently of the physical aspect, but there is something very like causality in the analytical aspect (logical entailment), juridical aspect (retribution) etc. Affordance [26, 27] is another example. While dependency is always on earlier aspects, analogy can be to any, anticipating later ones and retrocipating earlier ones.
Analogy provides linkages among the aspects that goes beyond the utilitarian notion of dependency, and is of many types. In particular, it is vital for knowledge representation and what Newell called 'aboutness' because it makes it possible for symbols to be 'about' things in any aspect - symbols of quantity, space, movement, ... love, commitment - in fact anything in the system's environment (kp:env). While the process of representation involves dependency on the pre-lingual aspects, analogy makes it meaningful, useful and robust rather than mere arbitrary attachment. See Fig. 3. 'Aboutness' is the very kernel meaning of the lingual aspect and, because, Dooyeweerd claimed, the kernels of aspects cannot be fully grasped by means of theoretical thought, but only by intuition, aboutness cannot be fully explained theoretically. This accounts for Newell's difficulty in doing so (kp:about).
The role of the lingual aspect is to capture any type of meaning in symbols. The content of the knowledge level medium results from this capture. The lingual aspect 'reaches out' to all the other aspects, and is the first one to do so. Therefore the content of the knowledge level medium can be much more diverse than that of lower levels, because their corresponding aspects lack this wide reach. This turns Newell's first 'surprise' (kp:variety) into something we would expect. A discussion of aspectual representation can be found in , who uses the term 'transduction'.
All three - irreducibility, dependency and analogy - are involved in design and implementation at any level (kp:des,kp:impln), which are to define the functioning in earlier aspects that is needed to support the functioning in the main aspect. As each aspect defines a distinct science, so each defines a technology (kp:tgy) in which implementation can occur. Dependency means that it will always be possible to find such definitions. From aspect to aspect, we can implement eventually as a physical artifact. Retrocipation means that we will usually be able to find implementations that are useful and meaningful. Irreducibility means that there is usually more than one possible definition, and in the reverse direction that a definition in an earlier aspect can often implement several things in later aspects (kp:m-n). It also affords considerable freedom to the implementors (kp:freedom) because they can abstract one aspect for attention and do not have to be hampered by detail from other aspects (kp:detail). Clouser  discussed two forms of abstraction: lower and higher. The latter is isolation of the aspect and holds dangers (of the kind we see in techno-centrism for instance) because Dooyeweerd held that the aspects 'resist' being separated. This might explain why  found that isolating the knowledge level from the symbol level can be troublesome.
Thus we have irreducibility, dependency and analogy. Notice the absence of the part-whole relationship. No aspect can be thought of as 'part' of another. Likewise, merely aggregating things of one aspect does not enter another (kp:agg). This is because the aspects are to do with Law rather than Entity, so the part-whole relationship is not as important as Immanence thinking makes it. However, it is discussed, especially by Hart , and thus components (kp:compn) can be supported but we do not discuss this in detail.
Because he presupposed Meaning as primary, Dooyeweerd saw behaviour in a different way. The tendency of an entity to behave in certain ways is found, not in the entity itself (Existence presupposition) but in Law (Meaning presupposition). Specifically, behaviour is functioning in one or more aspects. It is never mere activity but is always related to an aspect; it is always an outworking of the meaning of the aspect(s) in which the entity functions. So any description of behaviour at a level (kp:behav) is always a description of functioning in the relevant aspect.
Each aspect has its own distinct set of laws (e.g. those of syntax, semantics etc. for the lingual aspect) that govern or guide functioning in that aspect. Each aspect that corresponds with a level furnishes the laws of behaviour of that level (kp:lawb), so the principle of rationality (kp:por) can be seen as a subset of the laws of the formative aspect because of the importance of goals (kp:goal). In earlier aspects the laws are determinative (e.g. those of mathematics, geometry and physics) while in later ones they are non-determinative (e.g. the laws of linguistics, social interaction, aesthetics, justice) and can be seen as norms. The latter allow the entity some freedom of response, perhaps gradually increasing as shown in Fig. 4, though some thinkers believe it to be a step function.
Though, in later aspects, we have freedom both within the norms of the aspect and also to transgress the norms, they still pertain (and have repercussions) so we can predict behaviour (kp:pred). But prediction in normative aspects is no longer with absolute precision, but becomes rather an accounting for behaviour within each aspect. As discussed earlier, it is the laws of the ethical aspect of self-giving that enable us to discuss the princess's choice in meaningful ways.
Newell pointed out that behaviour at the knowledge level is according to a global principle (kp:glob), something outside the system and pertaining across all such systems, rather than a component's own local response to stimuli. We can see at once that aspectual laws can furnish such global principles. But Dooyeweerd would go further, to say that behaviour in all aspects, all levels, is likewise according to global principle. For example, physical material responds to global laws of physical aspect, structuring of symbols, to laws of formative aspect, etc. This means that Newell's second 'surprise' can be laid to rest. (There is evidence that Newell himself saw the possibility of global principles in all levels [KL:102].) The reason that behaviour of a physical component might appear to be its own self-generated response to stimuli, is because the laws of the physical aspect are determinative.
Most human behaviour is multi-aspectual, an amalgam of functioning in many (usually all) aspects. Usually the human is not explicitly aware of the aspects, but functions in them tacitly. For example, as a writer of this paper I function primarily in the lingual aspect, but I also function in earlier aspects such as the biotic (to keep alive while writing) and later aspects, such as the juridical (to give you, the reader, what is due to you). The user of a knowledge based system is also functioning in multiple aspects, including the lingual and all those on which the lingual depends (see above) and also the various aspects of the application, which are discussed later. Poor functioning in any one aspect can jeopardize the whole activity. This is why all levels are important (kp:imptnt) and we can agree with Newell when he said [KL:99] "Intelligent systems are not to be described exclusively in terms of the knowledge level."
Without Law no functioning is possible. So, to Dooyeweerd, Law is not a constraint (as it is to Western liberalism) but an enabling. Even normative laws are enablers - of meaningful, useful activity. But it should be noted that Dooyeweerd saw societal norms, for example, as very different from aspectual norms - as a culturally variable outworking of them - and in this way escaped the rigidity of some law-based frameworks of thinking. However, we should also note that Dooyeweerd's approach is the reverse of that taken by Carley and Newell . They start with all possible behaviour and narrow down in steps (rational behaviour, boundedly rational behaviour, .. to emotional behaviour). Dooyeweerd starts at zero and each aspect enables more and richer functioning. While the two may be logically equivalent, the emphasis is different, with Dooyeweerd interested in what is meaningful and useful. We can see that Dooyeweerd is able to bring together into a single framework the two poles of Determinism and Freedom. A slogan [unknown source] helps us understand his position:
Having obtained a general picture of functioning, we must now be more precise. Functioning in an aspect is of two types: subject and object functioning. Subject-functioning means that the entity, in and of itself, responds to the laws of the aspects in which it functions. Object-functioning, which we discuss below, is when the entity gains meaning within some other subject-functioning.
Dooyeweerd's approach to the subject-object relationship is novel, and Hart  discusses it in depth, showing how, for example, it integrates the various English language uses of the word 'subject': subject to law, subject in any sentence describing the functioning, and subjectivity as functioning in certain normative aspects. We omit much here.
Human beings can function as subject in all aspects, but non-human entities have a more limited range of aspects in which they can function as subject: animals up to the psycho-sensitive aspect (perhaps the analytical), plants up to the biotic aspect, and non-living things up to the physical aspect. This means that a computer, being a machine, can function as subject only as far as the physical aspect. Of itself and without any reference to human beings at all, it can only respond to the laws of the physical aspect (electric and magnetic fields, exerting force on a desk, etc.) That the physical aspect has determinative laws explains the reliability with which computer systems can function.
This, of course, seems to go against the assumption in artificial intelligence (and Newell's theory) of the equivalence of human and machine as knowledgeable actors. It also seems contrary to our everyday ascription of post-physical behaviour to the computer itself. We say "That building was designed by computer" when it was in fact an architect that designed it, using an intelligent CAD package.
The Dooyeweerdian approach can account for these by reference to object-functioning. Though the computer system cannot function as subject in these post-physical aspects, it can function as object. That is, the computer system can receive meaning from the aspects in which its user is functioning as subject, such as those above. This is no mere arbitrary or metaphorical attribution, but a substantive part of the matrix of meaning that is the application, something that the computer system plays an important part in. Therefore, to see the computer as 'only' a machine, though true in terms of subject-functioning, does not capture the full meaning of the situation. It may be that Dooyeweerd's treatment of the subject-object relationship can provide a framework for reassessing the unresolved debates about machine consciousness, intelligence, etc. because it frees us from having to define the Being (and attributes) of the machine, moving instead to the Meaning of the situation. But we do not discuss these here.
We can see the behaviour of the system described at each of Newell's levels as object-functioning that is validly attributed to the computer (in support of Newell against Dennett). At the bit level, the computer system is a psycho-sensitive object-functor, at the symbol level, an analytic object-functor, at the knowledge level (aboutness) a lingual object-functor, at the knowledge level (principle of rationality) a formative object-functor. (Note, we do not call it 'lingual etc. object' because of misleading connotations.) But this cannot occur without the subject-functor, whose meaning at each level-aspect is ascribed to the computer. This, of course, is identical to Newell's idea of observing and observed agents (kp:oo).
Wielinga, Sandberg and Schreiber  say it matters not whether knowledge resides in someone's head, in a book or a computer. Mentzas, Apostolu, Young and Abecher  say "knowledge exists in people, not technology." However, they can both be right, if we recognise that the former sees knowledge as object-functioning of the book or computer and the latter, as human subject-functioning; they are equivalent. This gives some support Newell's claim that all things that can be said to have knowledge are describable under this theory (kp:scope), except that Dooyeweerd distinct types of knowing (see later) might modify this.
But what of a computer system that is the observing agent, with knowledge and a representation of the observed agent (such as to create, describe or adapt it, including self-adapting systems that 'observe' themselves)? To say it 'observes' is to speak, again, not of subject-functioning but of object-functioning. The subject-functor is still human, though removed by one step, and attributes lingual object-functioning to the observing computer. The latter is physically subject-functioning in such a way that it might be interpreted by humans who observe it as describing, creating or adapting. Lenat and Brown  admitted that 60% of the reason 'Why AM and Eurisko appear to work' was because of human interpretation and guidance. Ultimately, this can be extended as many steps as needed, and the attributed object-functioning can be in any aspects.
To Dooyeweerd Meaning is primary, and this led him to an interesting concept of Existence and Entity that can throw more light on the existence of such things as bits, symbols and knowledge, and even on the computer system as a whole. Not only do Dooyeweerd's aspects designate ways in which things can be described and can behave, but, even more fundamentally, ways in which things can be (exist). To Dooyeweerd, Existence derives from Meaning, Entity from Law. His approach is able to address the philosophical issues of genesis, persistence and Heidegger's 'Dasein', and provides a way to discuss the essence of entities without carrying all the baggage of essentialism.
To 'exist' is to do so with respect to an aspect, specifically to function in that aspect as subject or object; Dooyeweerd spoke of aspects as "modes of being". For example, Nielsen's Fifth Symphony 'exists' by virtue of somebody functioning in the aesthetic aspect. A concept 'exists' by virtue of someone functioning in the analytic aspect. Popper's  World 3 entities 'exist' by virtue of functioning in the post-analytic aspects. And so on, for each of the troublesome 'existences' mentioned earlier. In a computer system, voltages and conductors, bits and registers, and symbols 'exist' by virtue of a computer user functioning in the relevant aspects. The 'existence' of knowledge, though, is more complex and discussed below. In fact, 'exists' is perhaps better replaced by 'gains its meaning' because impossible entities like a square circle can never exist, but do gain meaning within the spatial aspect.
Even so, the concepts 'exist' and 'entity' are ambiguous, so Dooyeweerd introduced technical terms: 'qualifying aspect', 'individuality structure', 'enkapsis' and 'enkaptic structural whole'. The qualifying aspect of something is the main aspect that gives it meaning; for example, documents are lingually qualified. (In fact, there are primary, secondary and tertiary qualifications , but we do not discuss them here.) An individuality structure is an 'entity' qualified by a single aspect, and normally tells only part of the story of a 'real' entity, which Dooyeweerd calls an enkaptic structural whole.
For example, (Dooyeweerd's example) Praxiteles' statue of Hermes and Dionysus 'exists' (or gains meaning) both as a block of marble (qualified by the physical aspect) and as a work of art (qualified by the aesthetic aspect). The statue of Praxiteles is not however just two individuality structures thrown together, but an integrated enkaptic structural whole in which both are necessary. Dooyeweerd discussed the relationship between necessary individuality structures, adopting the word 'enkapsis' from biology. (There are in fact several kinds of enkapsis , some of which can inform laws of composition of a level (kp:lawc).)
A computer system can be seen as an enkaptic structural whole with individuality structures qualified by several aspects that correspond to levels. This provides philosophic grounding for Newell's claim that at each level there 'is' a system (kp:sys), and that we can speak of the media and components at each level as though they are entities:
In particular, to say that a computer 'has knowledge' (e.g. in a knowledge base) is, in Dooyeweerdian terms, to say that its enkaptic structural whole involves a lingual individuality structure. The medium at each level (kp:med) about the materials seems to relate to Dooyeweerd's notion of semi-manufactured materials , but the details of this link have yet to worked out.
The problem of defining 'system' when our computer is connected to the Internet is ameliorated when we adopt a Dooyeweerdian perspective of Meaning and Law rather than Being and Entity. So it is no longer important to define what the system (or entity) is, but only what aspects are important - especially application purpose, below - and which laws are responded to. However Dooyeweerd treats things like environment and society - and so possibly Internet - in a different way from entities, but the relevance of this has yet to be explored.
Now we can turn our attention to the knowledge level in particular. The nature of knowledge has exercised the minds of philosophers over the years. Leibniz, for example, held that the individual represents the world to himself by means of subjectivity, Hegel added the notion of progress, and Habermas [29:313] saw knowledge as of three types: "information that increases the power of technical control, interpretation that make possible orientations of action within common traditions, and analyses that free consciousness from its dependence on hypostatized powers." Compared with this heritage, Newell's account of knowledge might appear rather simplified, but it does make a useful contribution, seeing knowledge, not as a commodity, but as the medium of a system, an active medium at that (kp:func).
While Dooyeweerd, who died in 1977, knew little of computers and knowledge representation, he too saw knowledge as active functioning that is knowing rather than the stuff that is 'knowledge' - in terms of "knowers, knowings and knowns" . Formative knowing involves goals and so includes the kind of knowledge that Newell employed in the principle of rationality (kp:por).
But Dooyeweerd went further than Newell did. Many assume there is one fundamental way of knowing (based either on reason or interpretation), and Newell seems to have been among them. Dooyeweerd held that each aspect provides a different way of knowing. Social knowing, for example, is radically different from analytical knowing, not just the latter employed for social purposes. Habermas' three types can be seen as corresponding to knowing in the formative, lingual and pistic aspects respectively. Anderson  seems open to distinct types of knowledge, suggesting that since we learn some operations at the algorithm level, these too can be called knowledge. This gives a multi-aspectual richness to Newell's idea of competence (kp:comp). However, Newell's statement that "knowledge is intimately linked with rationality" [KL:100] (kp:rat) is valid for all knowing that involves the formative aspect, namely all in post-formative aspects by virtue of inter-aspect dependency.
The aspectual view of knowledge enables us to extend the principle of rationality (kp:extn) by either replacing or supplementing it as the law of behaviour at the knowledge level with principles drawn from every aspect. The examples of additional principles given earlier (harmony, breakdowns, justice, self-giving) are drawn from Dooyeweerd's aspects. Our reference, above, to self-giving versus pride to predict the choice the princess makes, is the very kernel of Dooyeweerd's ethical aspect. Aspectual principles would provide a sharper instrument for prediction than rationality alone, and would have the added advantage of bringing Newell's theory more in line with post-rationalist perspectives in information systems. Newell's fourth 'surprise', that knowledge is an active medium, turns out to be one tip of the aspectual iceberg that Dooyeweerd brought to surface.
To understand the knowledge level we must understand the relationship between the observer, observed agent and environment (kp:oo, kp:env) - that is, taking the simple case, between user, computer system and environment. A person using a computer is engaged in two types of aspectual functioning: interaction with the computer and the tasks of the application. This gives us two sets of aspects, as shown in Fig. 5. We will speak only in terms of subject-functioning, for clarity.
True interaction arises from shared subject-functioning. So interaction between human and computer occurs at the physical aspect: force is applied by user to mouse, light radiates from the screen, etc. In later aspects (higher levels) the subject-functioning is only by the human being. Subject-functioning in the psycho-sensitive aspect (bit level), the light from screen is treated as signals sent to the brain for processing. Similarly for sound. Functioning in the analytic aspect (symbol level) the human interprets certain patterns as carrying symbolic meaning, while others do not (e.g. mere decoration). Functioning in the formative aspect the user forms concept structures. Functioning in the lingual aspect, the user attaches meaning (knowledge level 'aboutness' and also tacit level connotation).
So if the user recognises some knowledge in the computer, or perhaps manipulates it, it is because they are subject-functioning in the lingual aspect, and all the earlier ones on which it depends as we have seen several times above as corresponding with the levels. They are shown in the left-hand column in Fig. 5, Interaction. So, being able to describe the system at the knowledge level requires human interpretation (kp:interp) and is a stance taken by the interpreter (kp:stance). We can see again the need for an observing agent (kp:oo). If we were to ask the computer user what they were doing, they might answer, at the symbol level, "I am reading these numbers", or, at the knowledge level, "I'm working out what these numbers mean", and so on for each level.
But, supposing they were using the ELSIE system , they might answer "I'm setting the budget for our new office block, and trying to work out how long it would take to build it" (which are two of the tasks ELSIE was designed for). This answer is in terms of application task, the functioning of which happens to be in the economic and formative aspects. They might also use the INCA system  to write the contract they make with the construction company, which is functioning in the juridical aspect. As part of both tasks, they must think about spatial arrangements of the building, ground conditions, and the aesthetics of the design, bringing in the spatial, physical and aesthetic aspects. Thus, their behaviour in using the computer is a functioning in many aspects relevant to the application, which are shown on the right-hand column of Fig. 5, Application. A similar distinction may be seen in the two knowledge level slot schemes in Cyc .
The relationship between the two sets of aspects, shown by the fanning out that is similar to Fig. 3, is via the lingual aspect's 'aboutness', that is the knowledge the user interprets while functioning in the left-hand aspects. As discussed above, it 'reaches' for all other aspects via an analogical inter-aspect relationship and, in this case the aspects are those of the application (kp:appln) and the environment (kp:env).
Notice how there can be two functionings in an aspect, in this case in the physical and formative aspects, one concerned with using the computer itself and the other with the application. This can be confusing in a levels-based analysis, especially when the knowledge the computer system has is of symbols, mentioned earlier, e.g. for self-adaptive systems.
This allows us to understand more clearly the relationship between the two themes of the knowledge level, 'aboutness' and predicting behaviour, and the level-like Problem-space Computational Model.
In this way the three issues that seemed to confuse the original theory are now seen as inhabiting three distinct aspects, two of computer use and one of a specific application. So Fig. 2 can be recast with the help of Fig. 5 as Fig. 6.
We saw above that Newell's postulates of logical closure (kp:clos) and radical incompleteness of the knowledge level (kp:ri) generate problems, so it would be wise to replace them if we can. But if we do, we must ensure that the reasons Newell had for postulating them are recognised and, where appropriate, supported by other theory. Though Newell did not give explicit reasons for the postulates, we can infer three from what he wrote: to account for the generative nature of knowledge, to justify considering mixed-level systems, and to explain loss of determinacy.
First, that each aspect gives us a distinct way of knowing can explain the generative nature of knowledge (kp:genv) without recourse to logical closure. Representation is tied up with lingual knowing, but knowing in other aspects bypasses or transcends the lingual, so making it seem the agent knows more than has been represented. It might be no coincidence that Newell linked this to common sense; Dooyeweerd explained intuition and wisdom by reference to multi-aspectual knowing. Logic (part of the analytic aspect) is the mechanism proposed by Newell, but it is not the only one. To take Newell's example, when a person finds their way home at night they are functioning in the spatial and psycho-sensory aspects - and this might not involve any analytic reasoning at all, especially if we consider the capability of birds migrating. In terms of k1,2,3 of section 3.10, we can see that we need consider only k1 and k2. Logical closure, which focuses on a fictitious k3, is not needed and may be replaced.
Second, considering multiple aspects of use accounts for mixed systems, that take the agent's lower levels into account as well as knowledge level. In design and evaluation all relevant aspects in both columns of Fig. 5 should be considered because we saw above why all aspects are potentially important (kp:imptnt). But to predict behaviour in the way Newell seems to have meant, only the left-hand aspects need be considered.
Third, we must account for how determinacy is lost when moving from the (determinate) symbol level to the (non-determinate) knowledge level description (kp:lwrdet,kp:nondet). Dooyeweerd would see at least two sources of non-determinacy that are, respectively, of the right and left columns in Fig. 5:
Type (a) means, as mentioned above, that what our princess chooses is radically undetermined, because she is responding to the norm of self-giving love. Though predictions will not be precise, the analysis can be usefully informed by the aspectual norms, and also by the difference between what we might call valid and non-valid aspectual functioning, between that which exercises freedom within the norms of the aspect and that which transgresses the norms. To a computer, functioning as subject only in the physical aspect, type (a) does not apply.
Type (b) refers to the describer's freedom in the lingual aspect to interpret the physical subject-functioning of the computer in any way they wish as they compile descriptions of it. Freedom in description occurs at all levels, including the materials, so predictions at all levels are, thereby, non-determinate. This turns Newell's third 'surprise' on its head: he it took for granted that at all lower levels behaviour is determinate (kp:lwrdet) and was surprised that at the knowledge level it is not - but we find that at all levels there is non-determinacy, of type (b).
However, the notion of valid functioning helps us see that Newell's third 'surprise' contained an important element of truth. Almost any exercise of freedom in describing the materials, component and bit levels would be non-valid, resulting in error (e.g. describing a blank screen as not blank, wrongly stating a measured voltage, or ascribing bit 0 to 5 volts when the computer's designer meant it to be 1). The range of valid freedom widens as we ascend the levels, with a narrow range at the symbol level where, for example, programmers can interpret a character's bit pattern as a number. At the knowledge level not only do we have wider freedom of interpretation, but we usually gain a new dimension of non-determinacy, of type (a), because the knowledge is of the application aspects (kp:appln). So, if we assume valid descriptive (lingual) functioning, at lower levels the range of non-determinacy is exceedingly small when compared with the knowledge level.
So, for a full Dooyeweerdian treatment, we should re-express the claim of determinacy as follows. The computer system, of itself, functions as subject only in the determinate physical aspect, so its own behaviour is determinate. Its behaviour at the bit and symbol levels is object-functioning, as part of human subject-functioning in compiling descriptions. Hence, as long as the latter is valid, the object-functioning can seem determinate.
Therefore neither logical closure nor radical incompleteness are necessary to account for non-determinacy and the generative nature of knowledge. The notion of radical incompleteness (kp:ri) in the knowledge level is redefined to refer only to the lack of absolute precision in predicting the response of the agent to normative aspects (type (a)). A knowledge level prediction becomes an accounting for aspectual functioning according to the known laws of the aspects and to the notion of valid functioning. It should be able to fulfil this role without needing to be augmented by descriptions from other levels, so it can now, like other levels, fulfil its duty as a level (kp:compl). So the unfortunate disruption in the suite of levels is considerably reduced.
Dooyeweerdian philosophy seems to account for most of the key points, with the exception of the claimed need for logical closure and the radical incompleteness of the knowledge level, and even these can be replaced and extended rather than rejected. That a large number of key points are true of both levels and aspects:
makes it reasonable to treat Newell's levels as Dooyeweerd's aspects. The knowledge level in particular corresponds to the lingual aspect in the following key points:
We have also discussed how Dooyeweerd's philosophy can be used to address some of the problems we identified earlier as requiring philosophical treatment. Most fundamentally, to address the antinomy in the knowledge level, we abandon the 2,500 year old presupposition that Existence is self-dependent and the primary property of all that is, which inexorably leads to dualistic ground motives, such as that of Determinism-Freedom, and instead presuppose Meaning as primary, in which aspectual laws are a mix of determinative and normative. Having done that, the individual problems listed earlier can be addressed as follows:
How can Newell's theory of levels be made robust enough to serve the next twenty years, amid changing perspectives, technologies, fashions and needs? On reviewing and examining a number of problems with the theory, we found most were around the periphery. The heart of the theory - levels in general and knowledge level in particular - is sound, and thus likely to last.
Many of the problems required philosophical treatment, so we have sought a coherent philosophical underpinning for the theory. Neither rationalist-positivist nor interpretivist-perspectivist philosophies can fully support it, an inability we traced to presuppositions of the primacy of Existence and Entity that have underlain much Western thought for the last 2,500 years. In our era, this has resulted in a dualistic ground motive that presupposes that Determinism and Freedom are mutually exclusive. Yet Newell's theory seems to want to embrace both of these.
A new stream of philosophy, proposed initially by Herman Dooyeweerd , that presupposes Meaning and Law, deriving Existence and Entity therefrom, seems more promising. The central supporting idea is that Newell's levels correspond to certain aspects in Dooyeweerd's pluralistic suite, the knowledge level corresponding to the lingual aspect of symbolic meaning. The meanings of each level and corresponding aspect are almost identical, and their order is the same. Much that Newell claimed for the levels is what Dooyeweerd claimed for the aspects, and their notions of knowledge are very similar. Most gratifying of all, we have found that Dooyeweerd is able to support both the Determinism and Freedom sides of Newell's theory. The need for an interpreting observer (intentional stance) is not incompatible with holding "The computer knows X" to be more than in the eye of the beholder. (The latter comes from Dooyeweerd's interesting approach to aspectual functioning, entities and the subject-object relationship.) The degree to which Newell's theory is not only supported but enriched by Dooyeweerd's philosophy is remarkable, considering they came from totally different starting points (specifically secular artificial intelligence and Dutch neo-Calvinism).
Newell said [RKL] that "there is no way, given the current formulation [of his theory], to assess the degree of approximation [of descriptions] ... how systems could be made more intelligent ..." but we have suggested a way might be found via an understanding of the aspects and the notion of valid functioning. We have seen why the levels are as they are, why they are all needed, and why no intervening levels are possible. We have clarified some of complexity of the knowledge level, including Newell's later problem-space computational model, and resolved the most troublesome problems by sensitively expanding and redefining a few key points. The main emancipation required was from the positivism of the culture that surrounded Newell at the time, to allow for genuine non-determinacy and human freedom in both interpretation and normative, multi-aspectual behaviour.
To have generated a theory out of practice [KL:92] that has required so little modification when we sought a coherent single philosophical underpinning is a remarkable achievement by one man. It speaks of a lifetime of wisdom distilled in and filtered through that man. It speaks of creative thinking. It speaks of careful and painstaking work, as that one man thought through the "veritable jungle of opinions" [KL:92] about knowledge representation at the time, and produced a detailed scientific and technical proposal. But mostly, it speaks of courage. Allen Newell was working within a positivist and rationalist culture, that presupposed a completely determined cosmos and the detached and non-involved scientific observer, and that allowed no means for even discussing genuine non-determinacy or human interpretation. Knowledge had been reduced to symbols, psychology or sociology, and the theoretical attitude of thought reigned supreme. Yet Allen Newell knew that somehow knowledge should be treated as something in its own right, the messy type of knowledge from real life where princesses have to make impossible choices or people don't keep their promises, that non-determinacy and interpretation are facts of life, that the observer and observed must be considered part of a single dyad. This took courage, tenacity and brilliance, to see how these things could all fit together in a coherent theory.
These are the key points of Newell's theory, labelled by 'kp' for reference in the text, with their names, and an indication of how, or whether, Dooyeweerdian philosophy can support them. A short description of each can be found in appendix 1 in [CP]. They forms groups, of a main point, whose label and name are in bold text, with elaborating points associated with it.
|Support from Dooyeweerdian Philosophy
|Concerned with Suite of Levels
|Dooyeweerdian thinking gives dignity to intuitions, so that they can used in formulating theory, subject to certain precautions.
|Dooyeweerd starts, in the Prolegomena of his major work, not from theoretical analysis, but from everyday experience.
|Dooyeweerd's philosophy underpins or expands most of the key points.
|Each aspect defines a whole scientific area whose role it is to research those issues.
|Scope of Theory.
|Humans and machines are not equivalent in any aspect later than the physical. However, Newell's claim that all things that have 'knowledge' can be describable by his theory is supported if we speak of object- as well as subject-functioning.
|There is a plurality of aspects; each level corresponds to an aspect.
|The levels ML, CL, BL, SL, KL, TL
|Lingual aspect of representation plus preceding aspects.
|Each aspect can furnish us with a way of describing - concepts, vocabulary, etc.
|An aspectual description is valid if the thing described is involved in functioning in that aspect.
|Object of Description.
|Each valid aspectual description is of the same entity (enkaptic structural whole).
|Aspect can furnish a complete description.
|The developer functions fully in the left-hand aspects (giving different levels of design).
|Behaviour described at a level can be seen as functioning in the corresponding aspect.
|This makes it possible to predict, or rather account for, behaviour within an aspect.
|Dooyeweerd made an ontological claim for the aspects.
|No Intervening Levels.
|The aspects on which the lingual depends have no gaps between them.
|About Each Level
|Each level has ...
|Dooyeweerd's entity theory: individuality structures qualified by aspects within an enkaptic structural whole.
|Dooyeweerd's entity theory.
|The part-whole relationship.
|Laws of composition
|Laws of behaviour
|Each aspect has laws for functioning.
|Aspects are related in two ways: dependency and analogy.
|Each aspect depends on earlier aspects, which are necessary to it.
|Dooyeweerd's notion of a qualifying aspect.
|As each aspect defines a distinct science, so each defines a technology.
|Involves Dooyeweerd's inter-aspect relationships of dependency, analogy and irreducibility.
|The distinctness of the aspects accounts for the distinctness of the levels.
|The aspects are fundamentally irreducible.
|Each aspect has a distinct and irreducible kernel meaning.
|Aggregation of smaller things into larger (e.g. of atoms into molecules) does not take us to the next aspect. The part-whole relationship.
|Things meaningful in one aspect might not be in others, so could be invisible.
|Things meaningful in one aspect might not be in others, so could be inexplicable.
|There are many ways in which an aspect can relate to others, in each direction; inter-aspect relationships of dependency and analogy.
|Lower Level Detail.
|From kp:m-n, kp:irred.
|We can consider aspect independently of others, as lower or higher abstraction, but the latter can be dangerous.
|Poor functioning in any one aspect can jeopardize the whole activity.
|Concerning the Knowledge Level Itself
|The kernel meaning of the lingual aspect. Kernels of aspects cannot not be fully understood by theoretical thought but only intuitively.
|A description of a system is human subject-functioning.
|Represented knowledge is human subject-functioning in the lingual aspect, which involves various interpretation processes.
|Observer and Observed Agents.
|Without the human, the computer would function only physically. The observed agent is the computer system, object-functioning, and the observing agent is the human subject-functor.
|Each aspect provides a different way of knowing.
|Knowledge as Functioning.
|Dooyeweerd too saw knowledge as active functioning that is knowing rather than stuff that is knowledge.
|Knowing in various aspects transcends lingual representation in symbols.
|Process with Structure.
|Not specifically supported by Dooyeweerd, except by implication from his notion of what knowledge is.
|Logic is one aspectual generative mechanism; logical closure should be replaced by multi-aspectual knowing.
|The analogical inter-aspect relationship means the lingual 'reaches out' to the whole cosmos.
|The knowledge is of the aspects of the user's tasks which the use of the computer system aids.
|Lingual functioning cannot capture the complete meaning of anything.
|Variety from Content.
|The analogical 'reach' of the lingual aspect for the whole cosmos means huge variety.
|The non-absoluteness of the lingual aspect, as well as its wide reach, makes it arduous to express or represent meaning.
|Two types of non-determinacy: of functioning in the later application aspects and of making descriptions.
|Determinate Lower Levels.
|The physical aspect has determinative laws, the analytic and formative aspects have non-determinative laws. But, 'valid' interpretations of the computer's physical functioning seem determinate.
|As for kp:nondet.
|Aspectual laws pertain globally.
|This type of rationality is of the formative aspect, so any post-formative functioning or knowing involves it.
|Role of logic.
|Logic is of the analytic aspect, not the lingual.
|Principle of Rationality.
|One of the laws of Dooyeweerd's formative aspect.
|Goal-achievement is of the formative aspect.
|Knowledge level components.
|That the knowledge level components are actions, bodies of knowledge and goals can be supported, but more discussion is needed.
|PoR needs extending.
|We supplement or replace the principle of rationality with principles derived from all aspects.
|Not needed. Redefine it to cover genuine non-deterinacy in later aspects. Then knowledge level returns to completeness, though not absolute precision.
I would like to thank Aaron Sloman, my wife, Ruth, and my colleague, Andrew Faraday, for many useful comments on the paper. I thank God for giving us Allen Newell.
[KL] Newell A., The knowledge level, Artificial Intelligence, 18 (1982) 87-127.
[RKL] Newell A., (1993), Reflections on the Knowledge Level, Artificial Intelligence, 59 (1982) 31-38.
[UTC] Newell A., Unified Theories of Cognition, Harvard University Press, Cambridge, MA, (1990).
[CP] Basden A., The knowledge level - the first twenty years, (2002).
 Aamodt A., Nygård M., Different roles and mutual dependencies of data, information, and knowledge - An AI perspective on their integration, Data and Knowledge Engineering, 16 (3) (1995) 191-222.
 Anderson J.R., A theory of the origins of human knowledge, Artificial Intelligence, 40 (1-3) (1989) 313-351.
 Basden A., On the application of expert systems, International Journal of Man-Machine Studies, 19 (1983) 461-477.
 Basden A., Three levels of benefit in expert systems, Expert Systems, 11 (2) (1994) 99-107.
 Basden A., Engines of dialectic, Philosophia Reformata, 64 (1) (1999) 15-36.
 Basden A., Brown A.J., Tetlow S.D.A., Hibberd P.R., Design of a user interface for a knowledge refinement tool, International Journal of Human-Computer Studies, 45 (1996) 157-183.
 Basden A., Hibberd P.R., User interface issues raised by knowledge refinement, International Journal of Human-Computer Studies, 45 (1996) 135-155.
 Brachman R.J., Levesque H.J., Readings in Knowledge Representation, Morgan Kaufmann, Los Altos, CA, (1985).
 Brandon P.S., Basden A., Hamilton I., Stockley J., Expert Systems: Strategic Planning of Construction Projects, The Royal Institution of Chartered Surveyors, London, UK, (1988).
 Brentano F., Psychology from an Empirical Standpoint, Duncker and Humboldt, Leipzig, (1874).
 Brooks R.A., Intelligence without representation, Artificial Intelligence, 47 (1991) 139-59.
 Card S.K., Moran T.P., Newell A., The Psychology of Human-Computer Interaction, Lawrence Erlbaum Associates, Hillsdale, NJ, USA, (1983).
 Carley K., Newell A., The nature of the social agent, J. Math. Sociol., 19 (4) 221-62.
 Carroll J.M., Campbell R.L., (1989), Artifacts as Psychological Theories: The Case of Human-Computer Interaction, Behaviour and Information Technology, 8 (1994) 247-256.
 Chandrasekaran B., Allen Newell - Expert Interview, IEEE Expert, June 1993 (1993) 5-12.
 Checkland P., Systems Thinking, Systems Practice, Wiley, New York, (1981).
 Clancey W.J., Heuristic Classification, Artificial Intelligence, 27 (1985) 289-350.
 Clouser R., The Myth of Religious Neutrality; An Essay on the Hidden Role of Religious Belief in Theories, University of Notre Dame Press, (1991).
 de Raadt J.D.R., Information and Managerial Wisdom, Pocatello, Idaho, Paradigm, (1991).
 Dennett D.C., Brainstorms: Philosophical Essays on Mind and Psychology, Bradford, Montgomery, VT, (1978).
 Dietterich T.G., Learning at the Knowledge Level, Machine Learning, 1 (1986) 287-316.
 Domingo M., Sierra C., Knowledge level analysis of taxonomic domains, International Journal of Intelligent Systems, 12 (2) (1997) 105-135.
 Dooyeweerd H., A New Critique of Theoretical Thought, Vol. I-IV, Paideia Press (1975 edition), Ontario, (1955).
 Flood R.L., Jackson M.C., Critical Systems Thinking; directed readings, Wiley, Chichester, (1991).
 Gaines B.R., Knowledge management in societies of intelligent adaptive agents, Journal of Intelligent Information Systems, 9 (3) (1997) 277-298.
 Gibson, J.J., The theory of affordances, in: Shaw R., Bransford J. (eds.), Perceiving, Acting and Knowing, Erlbaum, Hillsdale, NJ, (1977).
 Greeno J., Gibson's Affordances, Psychological Review, 101 (1994) 336-342.
 Guarino N., Understanding, building and using ontologies, International Journal of Human-Computer Studies, 46 (1997) 293-310.
 Habermas J., Knowledge and Human Interests, tr. J.J. Shapiro, Heinemann, London, (1972).
 Habermas J., The Theory of Communicative Action; Volume One: Reason and the Rationalization of Society, tr. T. McCarthy, Polity Press, (1986).
 Hart H., Understanding Our World: An Integral Ontology, University Press of America, (1984).
 Heidegger M., Being and Time, tr. J. Macquarrie, E. Robinson, Harper and Row, New York, (1962).
 Hines J.G., Basden A., Experience with the use of computers to handle corrosion knowledge, Br. Corros. J., 21 (3) (1986) 151-156.
 Hintikka J., Knowledge and Belief, Cornell University Press, Ithaca, NY, (1962).
 Hirst G., Existence assumptions in knowledge representation, Artificial Intelligence, 49 (1991) 199-242.
 Hori M., Yoshida T., Domain-oriented library of scheduling methods: design principle and real-life application, International Journal of Human-Computer Studies, 49 (4) (1998) 601-626.
 Jennings N.R., On agent-based software engineering, Artificial Intelligence, 117 (2) (2000) 277-296.
 Kalsbeek L., Contours of a Christian Philosophy, Wedge Publishing Company, Toronto, Canada, (1975).
 Kammersgaard J., Four different perspectives on human-computer interaction, International Journal of Man-Machine Studies, 28 (1988) 343-362.
 Kripke S., Semantical considerations on modal logic, Acta Philosophica Fennica, 16 (1963) 83-94.
 Landauer T.K., The Trouble with Computers: Usefulness, Usability and Productivity, Bradford Books, MIT Press, Cambridge, MA, (1996).
 Lenat D.B., Brown J.S., Why AM and Eurisko appear to work, Artificial Intelligence, 23 (1984) 269-294.
 Lewis P.J., Information-Systems Development, Picador, (1994).
 Mentzas G., Apostolu D., Young D., Abecher A., Knowledge networking: a holistic solution for leveraging corporate knowledge, J. Knowledge Management, 5 (1) (2001) 94-101.
 Mingers J.C., Information and meaning: foundations for an intersubjective account, Information Systems Journal, 5 (1995) 295-306.
 Mitchell T.M., Learning and problem solving, Proc. Eighth International Joint Conference on Artificial Intelligence, (Morgan-Kaufmann), (1983) pp.1139-51.
 Peirce C.S., Reasoning and the Logic of Things>, Harvard University Press, Cambridge, MA, (1898).
 Popper K.R., Objective Knowledge, Oxford, (1972).
 Shneiderman B, Designing the User Interface: Strategies for Effective Human-Computer Interaction, Addison Wesley, (1992).
 Sloman A., Architectural requirements for human-like agents both natural and artificial, (What sort of machines can love?), Dautenhahn K, (ed.) Human Cognition and Social Agent Technology: Advances in Consciousness Research, John Benjamins, (2000) pp.163-95.
 Stafleu M.D., The idionomy of natural kinds and the biological concept of a species, Philosophia Reformata, 65 (2) (2000) 154-169.
 Stephens L.W., Chen YF, Principles for organizing semantic relations in large knowledge bases, IEEE T. Knowledge and Data Engineering, 8 (3) (1996) 492-6.
 Stockton F.R., A Chosen Few: Short Stories, Charles Scribner's Sons, New York, (1895).
 Tarnas R., The Passion of the Western Mind, Pimlico, Random House, (1991).
 van Heijst G., Schreiber A.T., Wielinga B.J., Using explicit ontologies in KBS development, International Journal of Human-Computer Studies, 46 (2-3) (1997) 183-292.
 Whitehead A.N., Adventures in Ideas, (1937).
 Wielinga B.J., Sandberg J., Schreiber G., Methods and techniques for knowledge management: what has knowledge engineering to offer?, Expert Systems with Applications, 13 (1) (1997) 73-84.
 Winograd T., Flores F., Understanding Computers and Cognition, Addison-Wesley, (1986).
Number of visitors to these pages: . Written on the Amiga and Protext.
Created: August 2002 Last updated: 22 February 2002 better sizing for pictures.