Knowledge Science:

A Pragmatic Approach to Research in Expert Systems1 T. R. Addis Computer Science Department, Univeristy of Reading, Whiteknights, Reading, Berks, RG6 2AY "God knows what the Truth is" Anon

Abstract
Underlying the notion of Expert Systems is the relationship between Knowledge and Technology. Technology has tended to force the notion of knowledge into specific categories (such as data and processes) that are inappropriate for modelling our understanding of the world. Our inability to resolve this tension between knowledge and technology has been one of the major reasons why Expert Systems are confined to a narrow band of application types. In order to provide an understanding into the relationship between knowledge and technology a pragmatic stance is taken (derived from the philosophies of C. S. Peirce) that relates directly our actions in the world to that of our knowledge about the world. A view of knowledge is presented that shows how the tension between knowledge and technology might be resolved. It is shown that representation and knowledge are distinguished as separate forms of knowledge and these forms must be brought together by the insight of an experienced designer in order to create an Expert System. A view is reached where the reason why our current systems are limited is made clear; new frontiers of research thus become obvious and definable.
This paper is based upon work that was supported by the Joint Research Councils' Initiative on Cognitive Science & Human Computer Interaction. Special Project Grant #9107137 (1991-1993). The work was done by Addis T. R. in conjunction with Dr D. C. Gooding and Ms J. J. Townsend (Science Studies Centre, University of Bath, Bath BA2 7AY).
1

1. Introduction
The study of knowledge is usually referred to as epistemology but this study is primarily concerned with the nature of people and their relationship to the world. In particular, the concern is centred on the justification of knowledge; knowledge that can be considered True. The difficulty in obtaining an absolute confident answer to this question as to whether a particular set of beliefs are True can be seen through the works of the French philosopher Rene Descartes (Sutcliffe 1968). He, having considered the possibility that even ones senses are suspect (we could be living in a dream, or hallucinating , or God could be playing a game), was reduced to starting his philosophies with the unconfirmed belief that God in His benevolence would not deliberately mislead us. Truth has thus been taken to be God's view of His world. From this underlying foothold formal representations of knowledge, such as logic and predicate calculus, have been created both of which in their turn underwrite the technology of the computer and the modern theories of artificial intelligence. Certainly ever since the philosopher Immanuel Kant established that knowledge depends upon our concepts it has been recognised that knowledge is inseparable from the mode of its representation. Knowledge engineers conceptualise expert knowledge in terms of their preferred method of elicitation and representation. Since there is a strong dependency between representation and knowledge, representations that developed within an engineering environment are normally biased towards particular machine paradigms (Shaw & Woodward 1990). Because of this bias such representations obscure certain aspects of knowledge, in particular, the different roles of different kinds of knowledge. It is therefore important that a general theory of knowledge representation that is cast for modern use should lie outside any particular representational paradigm.

2. A Taxonomic Approach.
The question "What is a knowledge representation?" was answered in part by Randal Davis et al (Davis et al 1993) and a summary of their deliberations is shown in Table 1. The important issue here is that it is a serious attempt at categorising knowledge by the role it has within a human/social/technological framework. However, the presumption is that knowledge remains only within the human domain and it is a system whose parts must necessarily remain within the world.

Role Surrogate Set of Ontological Commitments A Fragmentary Theory of Intelligent Reasoning A Medium for Effective computing

Description A Substitute for the thing itself. For reasoning about the world rather than taking action in it. In which terms should the world be thought about? Expressed in terms of: 1. the conception of reasoning. 2. the set of inferences sanctioned. 3. the set of inferences recommended. 1. An environment in which thinking is accomplished. 2. Guidance for organising information A language which says things about the world

A Medium for Human Expression Table 1. The Roles of Knowledge Representation: (Davis et al 1993) This restriction is explicitly made through the warning "A Knowledge Representation Is Not a Data Structure" where data structure may be some abstract representation scheme (such as a graph). However, knowledge is representable by a semantic net. The reason for this distinction seems to be based upon the idea that semantic nets have a 'semantics' where semantics refers to the topological constraints that come from what the net represents rather than the construction rules for the net. This loses a clear understanding of what the difference is between nets and graphs since the difference is identified through a set of unspecified implicit constraints; constraints that are only obtainable through human interpretation. This pushes the nature of knowledge back from whence it came; into the minds of people. If a proper study of knowledge is to be undertaken then the relationship between a representation and the world it represents must be investigated. Thus the important place that a user of a representation and its interpretation needs to be considered in depth. It is only then that an understanding of engineering Knowledge Systems can evolve. Addis theory expressed in part in this paper addresses this problem of engineering Knowledge Systems, including Machine Learning and Scientific Discovery (Addis 1985, 1989 & 1990). This theory is based on a taxonomy of knowledge which identifies the different types of knowledge called into play

when a system has to interact with the world to achieve some purpose. Thus, knowledge is classified according to its role in a system. Figure 1 shows the current status of this taxonomy, only part of which can be discussed here. Overlaid on this taxonomy are the roles, as identified by Davies et al, of the different categories of knowledge. All knowledge can act as a surrogate to the world. Tacit and declarative knowledge are primarily used via a representation as human expression and communication. Part of the declarative knowledge is the 'ontology' or taxonomy accepted by the participants of a conversation and the range of possible inference mechanisms that can be sanctioned are shown. Heuristics provide the guidance for problem solving within a representational scheme.
Surrogate
Knowledge

Expression
Tacit
Motor -skill Perception Facts Gestalt

Recommended
Declaritive Heuristic
Probability Corollarial

Sanctioned
Inference

Theorematic

Deduction Abduction Induction

Continuity Sample Prediction

Concepts
Abstraction

Retroduction Heuristic

Ontology
Taxonomic Distinctions Models Methods

Hypotheses Constraints
Theories

Finite

Infinite

Ordinal

Nominal

Order

Quasi-Random

Figure. 1. A Role Taxonomy of Knowledge. Knowledge can be divided into four broad categories. The theory implies that three of these categories: declarative heuristic inferential

must be present in any operational system, though their inclusion is not of course a sufficient condition for intelligence. The 'tacit' dimension remains outside any symbolic system but must be possessed by a human user for that system to be viable. Such knowledge is embodied partly in cognitive structures and partly in social ones; this is why so much of what people learn must be taught ostensibly, by example and by other people (Collins 1975 & 1990) 2.1 Tacit knowledge Tacit knowledge is a skill or understanding that can be demonstrated but cannot be represented as, say a list of rules or instructions. Typical examples of tacit knowledge are: riding a bike playing the piano most laboratory skills

problem solving (e.g. factors of a simple polynomial through inspection). Tacit knowledge may tentatively be divided into three distinct types: motor skill or kinaesthetic knowledge (the ability to interact physically with the world) perceptual knowledge (the ability to observe the world in an active sense) and gestalt (the ability to recognise patterns or 'situations' in the world).

Though many of these can be transferred (e.g. by instruction), tacit knowledge is personal in that its acquisition requires first-hand experience of the situations in which it is applied (Polanyi 1958); recent studies of scientific practice have emphasised the importance of personal knowledge (Collins 1975, Gooding 1990b, 1990c, 1992, 1993 & Shapin 1989). 2.2 Declarative knowledge Declarative knowledge is the main result of knowledge elicitation processes. It is representable knowledge. Much of a knowledge engineer's or analyst's work is concerned to identify the taxonomy (often referred to within artificial intelligence as the 'ontology') of an expert's domain. This involves

determining a set of distinctions made about the world, hypotheses or beliefs about that world, and facts taken to be true of that world. Facts . Facts are represented as propositions that ascribe specific properties to particular objects (or events). The ascription is based on observation. Thus, a fact is an instance of a concept and a concept is a generalisation from specific instances. Concepts can be used to imply new facts or to confirm known facts. Facts are 'known' observations, deductions or beliefs of or about the world. Taxonomic. The taxonomic class of concept consists of propositions that define general features, classes, categories, or types of objects or events, i.e. it is concerned with the distinctions made in a domain. These propositions are identified by the nature of the knowledge in that they are 'true' by convention or by definition. For example, the truth of the statement that "a bachelor is an unmarried man" depends upon a convention that the verbal token bachelor should represent the class of persons who are male and unmarried. Hypotheses. Hypotheses, on the other hand, are propositions that express constraints, laws or rules about the world rather than our modes of representing the world. Taking a pragmatic, Peirceian view (Peirce 1958 & 1966), "hypotheses are true if they are useful; they are useful if they make the world "a less surprising place". In conjunction with inference, hypotheses make predictions about the world so as to reduce the effective information content of incoming messages from or about the world. Without hypothesis formation, it would not be possible to develop new concepts about the world: learning could not happen. Table 2 summarises the distinctions between different types of declarative knowledge. Note that the role of Truth here is to characterise an abstraction of the world that is person or people centred. Thus the set of facts, distinctions and hypotheses declare a view of the world that is in part influenced by the world itself (observations), part by the individual (beliefs), part by society (conventions) and part by purpose (convenience). Truth reduces to a parameter that really has a role only for deductive inference. Types of Declarative Knowledge Fact Distinction (Taxonomic) Hypothesis Truth of the Knowledge Dependent upon:Observation, Deduction or Belief Convention or Definition Convenience

Table 2. The Role of Truth in Declarative Knowledge Hypotheses may be either theories or models. Although these are often represented visually (i.e., as visual hypotheses (Gooding 1992, Thagard & Hardy 1992, Trumpler 1992)) we are in the first instance only concerned with the declarative, verbally articulable forms of hypotheses. However, we will wish later to extend the notion of a hypothesis to form part of a mechanism (a machine, or a physical structure or a natural system). Models, in conjunction with an inference system and heuristic knowledge, make predictions about a limited part of the world and are derived from theories that encompass a greater view than any model. Associated with a model is a set of inference mechanisms so that the model may be activated (driven). Many different models may be derived from a single theory. Thus, a set of models is the extension of a theory in the same way that a set of facts is an extension of a concept. An extension is a relationship between a set of symbols and a symbol which is not itself a member of the set. The meaning of the symbol is the set. Since every symbol of the set can itself have an extension, the relationship of concept to a set of facts (or theory to a set of models) is relative and not absolute (Addis 1990). 2.3 Heuristic knowledge Heuristic knowledge is that knowledge which indicates how a deductive process should be performed. Given a number of facts and hypotheses many potential paths of deduction are possible. The heuristic knowledge uses extralogical information to guide the deductive process towards 'useful' results. What is useful depends upon the purpose of the system. Peirce pointed out there are three types of deduction . Since there is only one formal process of deduction based upon 'modus ponens' (although there are many mechanisms such as resolution that can achieve this result) the distinctions he identified can be related to the different classes of heuristics identified in Figure 1. Exactly how these may be identified within knowledge systems is yet to be understood. 2.4 Inferential Knowledge Inference can be divided into two styles:

Open: 'Open' inference does not necessarily have identifying formal mechanisms associated with it. In general, open inference is related to the creation or the construction of insights.

Closed: Closed inference identifies the process of selection from pre-formed or pre-determined insights. The problem with most AI is that it simulates intelligence through inference that appears open (capable of learning) but which is in fact closed. Closed inference identifies a style of inference usually found in both AI mechanisms and many established skills (such as navigation, engineering and mathematics). Inference, as Peirce observed, has three distinct forms: abduction, induction and deduction: Abduction. In its most general form, abduction is the process behind insight. It is involved in three classes of activity: Retroduction: Open retroduction is the creation of a new hypothesis and closed retroduction is the selection of a hypothesis from a pre-defined set. Closed retroduction is often referred to as 'abduction' in the literature. It depends upon the notion of reverse implication. Thus, if A->B and we know B then A is a possible cause. However, this inference depends upon the preexistence of A->B. Abstraction: Open abstraction is the process of creating or observing new taxonomic concepts and closed abstraction is the process of selecting taxonomic concepts from a pre-defined set. Heuristic: Heuristic abduction is the insight that creates or selects the process (the heuristic) on how to solve a problem. The heuristic informs the inference process as to how to continue with deduction. Whereas heuristic knowledge selects a path of reasoning, the heuristic abductive inference proposes how such a decision should be made. Induction. Open induction, taken in the context of the other forms of inference defined here, is the process of validating a hypothesis. Validation involves induction that is not enumerative, that is, does not involve generalising from a set of particulars. Induction requires the pre-existence of a hypothesis, a

deductive procedure and a set of criteria. The categories of criteria are shown in Figure 1. A common definition of induction, based upon the principle of generalisation from a given set of instances, in practice combines the notions of validation and retroduction (Strawson 1952). 'Induction' in this combined sense is the process of reasoning where the conclusions are not entailed by the premise (i.e. truth is not guaranteed to be preserved). Thus statements like: "He's been travelling for 24 hours, so he'll be very tired" can be explained by suppressed premises (undeclared hypotheses). Such suppressed premises allow us to reconstruct the 'inductive' inference as a form of deductive inference. However, the suppressed premises do not appear as conclusions to deductive arguments based on particular instances (examples drawn from the extensions of the premises). 'Induction', as held by this common view, thus involves the generation of a suppressed premise (retroduction) and its validation (induction as defined here). On the view of Addis, Gooding and Townsend (Addis et al 1993) validation involves assessing the utility or viability of a hypothesis, given a particular purpose. Closed induction validates by comparing the inferred results with an established model (e.g. the null hypothesis in statistics). Induction takes on three categories of criteria: sample, continuity and prediction. For example the 'sample' criteria is concerned with testing a hypothesis against a set of examples (which may be finite or infinite). The validation of a hypothesis through its ability to make predictions (that is to reduce the apparent entropy of the given set with respect to a set of classes) is well defined for a finite set. Validation of an infinite set given only a finite set of examples must make some assumptions on the stability of the properties of the set. Deduction. Deduction on the other hand is valid through formal proof. It is the process of inferring conclusions that must follow from the premises; it preserves Truth. Closed deduction uses known (i.e. pre-determined) solutions.

3. Intelligent Inference

Within this framework we define as 'intelligent' any system that has the capacity to use all three forms of inference together (NB. Abstraction and Retroduction are classed together as Abduction in this definition2). Intelligent processes typically keep these forms in tension. The dynamics are more complex than the logical dialectic of 'generation and 'testing'. As studies of knowledgecreation processes in science show, its character is continually changing (Gooding 1992, Pickering 1989, Tweney 1985, Tweney & Gooding 1992). The main cycle of activity may be described as follows: Abstraction of relevant features ensures that a germane hypothesis can be retroduced. Retroduction attempts to create a hypothesis (or set of hypotheses) that is framed in terms of the abstracted features and fits given knowledge and understanding. The creation process is dependent upon some mechanism (meta-concept?) that is influenced by experience. In this way only feasible hypotheses are generated. Deduction from this hypothesis will provide the essential information for induction. Induction validates the deduced results according to a set of criteria that depends upon purpose. The result of induction influences further abstraction, and so on.

It is not clear to the author if the process of Abstraction is necessary for some kind of 'pure' intelligence where the elements of the world are given or predefined.

Abstraction

Retroduction

Induction

Deduction

Figure 2. The Cycle of Intelligence and the Interaction between the different types of inferences Figure 2 illustrates the process of intelligence described above. The creation and validation of hypothesis are performed by interacting and cooperating abstraction, retroductive, deductive and inductive inferences. As suggested this process in its simplest form is similar to the simple 'generate and test' procedure. However, the process is more complex than this simple cycle in that the results at each stage influence the way in which each element in the cycle behaves. There is a "tension" among the three inferences and this "tension" provides feedback data from one inference to another in order to improve the quality and credibility of a potential hypothesis. The black solid circle in Figure 2 indicates the flow of information (thick arrow) between the four inferential mechanisms. The tension is created when the four inferential mechanisms cooperate to formulate a viable hypothesis. Communications between the mechanisms involves a cycle of abstraction, retroduction, deduction and induction and the feeding back of information which will ensure a hypothesis is applicable. Thus features from the world are abstracted that serve the overall purpose of the system and lead towards viable hypotheses. Hypotheses are abduced that will, in their turn, serve the criteria of induction and deduction will form conclusions that can be validated and useful (e.g. solve the problem). Abstraction is unique in that it depends to a large extent on close activity with the world; it is sense based (i.e. seeing, touching, hearing, tasting and smelling as well as using the propriaceptors) For a given set of facts (e.g. a sequence of numbers) it is the role of retroduction to create a reasonable hypothesis for those facts (e.g. the sequence). Deduction exercises the hypotheses and returns a prediction of a new fact (e.g.

the next number in the series). The validation of the retroduced hypothesis is the function of induction that ensures that the hypothesis is suitable for the purpose (e.g. the prediction is correct, the calculation was not to complex, it fits all or most of the facts and the form of the hypothesis is simple). The interaction of the three forms of inference explains why, given that many possible hypotheses can be generated, we tend to generate only a small set for evaluation. In the Faraday simulations done by Gooding and Addis (Addis et al 1990, 1992, Gooding 1990a, 1990b, 1992) one example of this appears as the Construe-Experiment-Clarify (C-E-C) cycle. It should be possible to identify different reasoning styles (or different qualities of intelligence) according to the mix or range of different kinds of inference.

4. Knowledge Acquisition
The knowledge engineers task is to represent and model expert knowledge. This involves negotiating a path from the informal and prearticulate state of an expert's (or experts') knowledge towards a formal model of it. Each stage of this process has its own representation scheme and associated techniques. It was argued by Addis et al (Addis et al 1993) that because a functional database language (FDL) explicitly represents all the different declarative roles of knowledge it will therefore provide a formal representation that is best suited for an unbiased elicitation method. They then argued that a graphic approach to formal modelling exploits the advantages of visualisation, and that a graphic approach offers a median way between the extremes of rules-oriented and dataintensive approaches. By combining the two argued points in a Visual Functional Program (in a language called CLARITY3) they provide the justification that VFP is an ideal environment for knowledge acquisition. 4.1 Abstracting to a Representation. An emphasis on the linguistic form in most knowledge representation defers to the traditional philosophical bias towards propositional knowledge. In practice, most of the conceptual modelling representations involve an essential pictorial or diagrammatic representation, even though these may be defined in terms of a language. Systemic networks, semantic nets, SFD graphs, KADS diagrams and repertory grids express conceptual models within a two

A beta test version of CLARITY is available through the author.

dimensional diagram. Relational analysis uses a diagrammatic scheme to show the elementary items (attributes, objects and entities) that are
Target system

interaction Mental models Elicitation procedures Intermediate knowledge bases communication Analysis procedures

introspection Conceptual models

models of the conceptual models

Visual & Functional

CLARITY

Computer knowledge bases Implimentation procedures

formalization

Modeling requirements

Cognitive processes

Acquisition procedures

Knowledge representation

PSYCHOLOGICAL THEORIES

KNOWLEDGE ACQUISITION TOOLS

Figure 3. Abstracting to a representation4 recognised by a business. However, many diagrams are too informal to translate directly to a formal structure that can be implemented as a program. For these -as where the target system (implementation) forces a representational bias. -- an extra stage in modelling is required. Figure 3 shows the processes linking elicitation, modelling and formal representation (modified from reference Shaw and Woodward 1990). The nonverbal representations typically used in moving from conceptual models to formalised, implementable ones, are not shown. It is well known to AI workers that sketches and diagrams are good guides to implementation. Nevertheless, accounts assume that all diagrammatic modelling reduces to linguistic representation. Thus an account by Shaw and

The processes linking elicitation, modelling and formal representation (modified from reference Shaw and Woodward. The non-verbal representations typically used in moving from conceptual models to formalised, implementable ones, are not shown.

Woodward, in which diagrammatic representations are not even identified as part of the process , shows how quickly visualisations are replaced by linguistic representations. They therefore underestimate the importance of diagrams to communication and conceptualisation during the elicitation process (Figure 3). Of course the machine representations sought by knowledge engineers -- like Maxwells famous field equations -- must be symbolic in form. However, this fact does not require us to assume that all representations reduce to linguistic ones. Images were just as important to Maxwell as they had been to Faraday (Wise 1979). Many modes of scientific investigation are irreducibly graphical (Griesemer 1991 & O'Hara 1992). Formal representation is essential to computational representation, nevertheless it may prematurely displace informal, diagrammatic working during the elicitation process. (Shaw & Woodward's summary 1990 of the process (pp. 189-90) suggests that this is the case). Addis et al give the diagrammatic representations used in knowledge elicitation a more prominent and more enduring role by showing that it is possible to combine the accessibility, flexibility and exploratory capacity of diagrams with the disciplines of formal representation. It is possible to have your cake and eat it because it is possible to generate program code directly from diagrams. Moreover, since a functional database language (FDL) is used, it is possible to model at any level of abstraction, combining different levels of complexity in the same representation. Further advantages of this approach emerge from a consideration of the recent history of hardware and systems software design.

5. The bifurcation of data and processing.

Influenced by earlier conceptions that the world can be split into two categories, data (information) and processing (action), computer technology has dictated that all knowledge (including models of skills and the world) should be divided between storage (data) and processing. Work on rule extraction from data using 'inductive' techniques has demonstrated an equivalence between these two forms of knowledge (Quinlan 1979). Models of skills (such as the skilful playing of a chess endgame) can be represented either as a large database processed by a simple pattern matching algorithm or as a set of rules that is processed by a deductive inference system.

Processor Imperative Program

Nassi Schniederman Charts Faith
Object Orientated Analysis

Knowledge System Declaritive Program

Clarity
Function Chart

Data Flow

World

Files

Store

Abstract to Storage

Figure 4 Abstracting to a representation5 Similarly, software engineering has tended towards design methods that emphasise either the data or the process form of representation. Systems analysis, in particular, is oriented toward database designs. These use graded methods that first consider the domain directly and then move the abstraction of the domain towards a mass data storage solution. The steps involved are: Relational Normalisation, Conceptual Model, Logical Model and Physical Model (see Figure 4). Other methods, largely independent of database analysis, develop the associated software (process). These methods include: Function Charts, Data Flow and Nassi Schniederman Charts (Figure 4). The advantage of these approaches is that each can provide a method; an explicit procedure that ensures a reliable design that can be justified. However, these methods are appropriate only for the imperative programming of data. Because of their influence there is a tendency for knowledge analysts to bifurcate knowledge into taxonomic structures (suitable for storage) and processes (appropriate for imperative programming), as a matter of course.

The diagram illustrates the emphasis of the two main software engineering methods on process (vertical axis) and information (horizontal axis). Any model of knowledge can be represented at some level of abstraction as a point in this space. The scales of abstraction indicate the order that each method (enclosed in an oval) is used to implement a system.

10 Mathematical Model Processor Store

Knowledge

Processor 4

Store Processor Store

Knowledge Knowledge Area of Conception Database System 0 5 Data Size (10n ) 10

Figure 5. The interchange of Data with Rules (D. Michie, 1979) 6 However, as D. Michie pointed out, people do not naturally conceive their own knowledge in either of these extreme forms of data and processing. Representation of peoples knowledge is best understood in terms of some intermediate form that combines data and rule. Many practical skills use such a mixture: navigators, statisticians, architects and electronic engineers all use tables as well as rules. Michie called this intermediate form "the human window" (Michie 1979). It is the shaded area in Figure 5. Interactive systems -- whether used by experts or by novices -- should operate within these limits. The diagram (Figure 5) suggests that knowledge can be considered as something apart from its representation. The same knowledge can thus be described in different forms in the same way that music can be represented as the physical undulations on a disc, or the magnetic orientations on tape, or the musical notations on paper. A knowledge system consists of a balance between a pure store and a processor. The role of object oriented analysis is to keep the balance between data and processes for any particular knowledge source. This balance ensures
The inverse relationship between process and data for a single source of knowledge. The curved line indicates the many different mixes of data and process. The spheres denote a single source of knowledge represented in different ways (Addis , 1980). Michies 'Human Window' is labelled as the 'Area of Conception'.
6

that the designer can conceive of his model more easily, in the form of the object. 'Frames' -- the original conception that led to object analysis -- has been developed empirically. The object form therefore lacks the coherence and completeness that might be expected, had it been derived from a formal theory of knowledge representation.

6. Conclusions
Knowledge of the world is a human and social affair. However, it does have identifiable elements that allow us to create hypotheses, represent knowledge and discuss the nature of intelligence. Given such a study it is possible, within the human framework, to attribute knowledge to inanimate objects; objects such as computers and other machines. This is because we choose to interpret its form in a way that corresponds to our view of the world. What has not been discussed in this short paper, and is considered in more detail in 'Knowledge for Design' (Addis, 1990), is the role and nature of the representation schemes themselves. Briefly, representations exist in the world, have constraints and exist in the domain of knowledge. What a designer of a system has to do is to match a representation and its constraints to the knowledge of the domain under consideration. The most powerful representation schemes are those that have a long history and have evolved strong inference procedures. Mathematics, in all its forms, is an example. It is therefore an important strategy to represent the world in a mathematical form for then it is possible to call upon centuries of thinking to support any conclusions that may be drawn; it is the route to successful scientific advance. What is not known is the key to systems that we could accept as truly intelligent. This key is dependent upon how to achieve an 'open' inference system. One clue to this achievement is the interaction of an intelligent and purposeful system with the world; particularly at the level of abstraction. The world, after all, is the only unbounded system we have and it is through our interaction with the world that new concepts and novel systems of concepts emerge.

7. References
Addis T. R., (1985) Designing Knowledge-Based Systems, Originally Kogan Page, now Chapman & Hall. published October. Hardback ISBN 0 85038 859 7. Soft back ISBN 1 85091 251 3., . Addis T. R., (1989) 'The Science of Knowledge: a research programme for knowledge engineering', Proceedings of the Third European Workshop on Knowledge Acquisition for Knowledge-Based Systems (EKAW' 89), Paris, June, . Addis T. R., (1990) 'Knowledge for Design', Knowledge Acquisition, Vol. 2, pp 95-105, . Addis T. R., Gooding D. C. and Townsend J. J. (1991) 'Modelling Faraday's Discovery of the Electric Motor: An investigation of the application of a functional database language'. Published in the proceedings of the Fifth European Knowledge Acquisition for Knowledge-Based Systems Workshop, Crieff Hydro, Scotland, 20-24 May. Addis T. R., Gooding D. C. and Townsend J. J. (1992) 'A Principle-Based Engineering Environment'. Presented at the IEE Colloquium. Digest No 1992/191, pp 2/1 - 2/10, (November). Addis T. R., Gooding D. C. and Townsend J. J. (1993). 'Knowledge Acquisition with Visual Functional Programming'. Knowledge Acquisition for KnowledgeBased Systems, 7th European Workshop, EKAW '93, Toulouse and Caylus, France. September 6-10. Pub Springer-Verlag, ISBN 3-540-57253-8, ISBN 0-38757253-8. Lecture Notes in AI 723. Addis, T. R., Gooding D. C. and Townsend, J. J. (1992) 'Knowledge Design From Natural Intelligence: An Interactive Representation Of Faraday's Experimental Researches', Poster, Joint Councils Initiative Workshop, Imperial College [March ]. Addis, T. R., Gooding D. C., and Townsend, J. J. (1990) 'A functional description of Faraday's discovery of the electric motor', Working Paper, Science Studies Centre, University of Bath, . Collins, H. M., (1975) 'The TEA Set: Tacit knowledge and scientific Networks', Science Studies 4: pp.165-185, . Collins, H. M., (1990) Artificial Experts, Cambridge, MA.: MIT Press,. Davis R., Shrobe H. & Szolovits P. (1993) 'What is a Knowledge Representation' AI Magazine, Vol. 14, No 1, pp 17-33, Spring. Gooding D. C., Addis T. R. and Townsend J. J. (1992) 'Knowledge Design from Natural Intelligence: An Interactive Representation of Faraday's Experimental Researches. Poster, Joint Councils Initiative Workshop, Imperial College, (March). Gooding, D. (1990a)'Mapping experiment as a learning process', Science, Technology and Human Values, 15, pp. 165-201. Gooding, D. (1990b) Experiment and the Making of Meaning, Dordrecht: Kluwer Academic, .

Gooding, D. (1990c). Bath 3: Rediscovering Skill in Science, Technology and Medicine, Bath: Science Studies Centre, Gooding, D. (1992) 'Putting Agency back into Experiment', in A. Pickering, ed., Science as Practice and Culture, Chicago: Chicago University Press, pp. 65-112. Gooding, D. (1992a) 'Simulating Real Experiment', International Workshop on The Replication of Historical Experiments, Oldenburg, August. Gooding, D. (1992b) 'Mathematics and method in Faraday's experiments', Physics., XXIX, pp. 121-47 . Gooding, D. (1993) 'Visualisation in and of science', Joint HSS-BSHS-SHOT Conference, Toronto, July 1992 and Princeton University Workshop on Visualisation in Science, February Gooding, D. (1993) 'What is experimental about thought experiments?, in Hull, Forbes and Okruhlik eds. PSA 1992, vol.2., 1993, in press. Griesemer, J. R. (1991) Must scientific diagrams be eliminable? The case of path analysis, Biology and Philosophy, 6: 155-180, . Michie D. (1979) Problems of the 'Human Window'. Presented at the AISB Summer School on Expert Systems, Edinburgh, July . Nersessian, N. J. (1984) Faraday to Einstein: Constructing Meaning in Scientific Theories, Dordrecht: Nijhoff, . Nersessian, N. J. (1992) How do scientists think? Capturing the dynamics of conceptual change in science, in R. Giere, ed., Cognitive Models of Science, Minneapolis: University of Minnesota Press, , pp. 3-44. OHara, R. J. (1992) Telling the Tree: narrative representation and the study of evolutionary history, Biology and Philosophy, 7: 135-160, . Peirce C. S. (1958) Science and Philosophy: collected papers of Charles S. Peirce , v.7, Harvard University Press, . Peirce, C. S. (1966) Charles Peirce: Selected Writings, edited by P. P. Wiener, New York: Dover, . Pickering, A. (1989) 'Living in the material world', in Gooding, Pinch and Schaffer, eds. , pp. 275-297. Polanyi, M. (1958) Personal Knowledge: Towards a Post-Critical Philosophy, New York: Harper & Row, . Quinlan J. R. (1979) Discovering Rules by Induction from Large Collections of Examples. Expert Systems in the Micro Electronic Age, Ed D. Michie, pub Edinburgh University Press, , .pp. 168 - 201. Shapin, S. The invisible Technician, Science, 77: 554-63, (1989). Shaw L. G. & Woodward J. B. (1990) Modelling Expert Knowledge, Knowledge Acquisition ,vol. 2, pp179-206 . Stender J., Addis T. R. and Spenceley S. E. (1993). 'Principle-Based Engineering and Economic Modelling'. Parallel Genetic Algorithms, ed. Stender J., IOS Press, Amsterdam, ISSN: 0922-6389. pp 117 - 128.

Sutcliffe F. E. (1968) "Descartes: Discourse on Method and Meditations", Penguin Classics, ISBN 0 14 0444.206 5. Thagard H. & Hardy S. (1992). Visual Thinking in the Development of Dalton's Atomic Theory, Proc' 9th Canadian Conference on AI, Vancouver Trumpler, M. (1992) Converging Images: Techniques of Intervention and Forms of Representation of Sodium Channel Proteins in Nerve Cell Membranes, paper for Princeton History of Science Workshop Series on Visualisation in the Sciences, October . Tweney, R. D. & Gooding, D. (1992) Introduction, Michael Faradays Chemical Notes, Hints, Suggestions and Objects of Pursuit of 1822, London: IEE, . Tweney, R.D. (1985) Faraday's discovery of induction: a cognitive approach, in Gooding and James, eds. , pp. 189-209. Wise, M. N. (1979) 'The mutual embrace of electricity and magnetism', Science, 203: 1310-18, .