Process models

Processes are ?ften represented by physical analogues or by mathematical equations. An example of the former is the use of a model of an aircraft in a wind tunnel. Measurements can be made on this and the likely effects on a full-sized plane of similar composition and design can be assessed. An example of a mathematical model is the use of probability functions to represent what happens in a queue over a period of time with particular patterns of service and arrivals.

With the advent of computers a new medium for representing processes became available. A computer program, when it is executed, progresses through a sequence of states. If these states have some correspondence with the states of a process then a useful model may be developed. This is the basis of simulation which can be regarded as a dynamic model of a process that 'unfolds over time' (Rothenberg, 1 989, p82). Computers are not essential for simulation, of course. The changes in a model could be recorded on paper. Often, however, there are complex or large numbers of calculations to be carried out which makes anything but a computer model infeasible.

In all the examples described so far there has been a requirement to measure quantitative aspects of a model whether it is the level of vibrations in a model aircraft or the number of elements in a simulated queue. If the model is to be used for teaching purposes then numerical calculations may be le�s significant than q ualita tive considerations.

Qualitative models are of great interest in artificial intelligence since, as observed in the previous chapter, both experts and novices seem to rely on descriptions to guide their problem solving. Clancey notes that such models describe systems in terms of 'causal, compositional or subtypical relationships among objects and events' (Clancey, 1 989, p l O). From this viewpoint, knowledge bases are predominantly qualitative whether they represent expertise at a deep or shallow level. More information on the contrast between these two approaches to knowledge representation can be found in papers by Chandrasekaran and Mittal ( 1 983), Hart ( 1 982) and Michie ( 1 982). B asically, shallow knowledge bases represent observable rules and heuristics that experts employ. In contrast, deep methods attempt to represent a fuller knowledge of the subject area and the reasoning that the expert uses.

For our purposes, Clancey's categorization of process models is a useful one and is reproduced as Figure 3.5. Classification models are often at a shallow level and seek to provide some pattern with which the original can be compared. The heuristic classification provided by the rules in such systems as MYCIN (Shortliffe, 1 976) and MUD (Kahn and McDermott, 1 984) are typical examples. Frames and scripts are tools that are often used to aid classification.

Types of process models

Classification Simulation

(critical patterns characterizing (cause-effect descriptions) objects and events as processes)

Example: disease prototypes

Behavioural Functional/Structural

(transitions among prototype state descriptions) Example: Causal network of

pathophysiologic states

(composition of structural and functional modules yielding behaviour)

Example: Hierarchical description of body organs and systems

The distinction between behavioural and functional/structural simulation is important and has been a source of confusion in the area of expert systems . For example, B obrow ( 1 984) in a special issue on qualitative reasoning considers only functional/structural models. It is useful to elaborate further on these two different kinds of simulation. Broadly, the distinction is between modelling obse rvable characteristics of a system taken as a whole, and modelling the individual components, describing what they do and how they interact.

If there is a detailed model of how the observables in a system are caused (for example, NEOMYCIN gives a description of how diseases exhibit particular symptoms) then, by applying a reasoning procedure to this model for a specific case, a causal, situation­ specific model (Clancey, 1 988c) of the process can be produced. The behaviour of the system in particular circumstances is being simulated. This simulation, the kind that occurs in many deep expert systems, can be effective but is very limited. There is no attempt to model the behaviour of systems in any general way (for example, the physiology of a healthy patient) or to simulate the functioning of individual components (such as the heart or lungs in a medical model) . Instead there is an emphasis on manifestations of particular kinds of phenomena and their causes. Often the model is expressed as system states linked in a causal associational network.

An example of a system that uses this idea is CIRCSIM-TUTOR (Khuwaj a et al. , 1 992). This is an ITS for teaching students the principles of the baroreceptor reflex, the part of the cardiovascular system that controls blood pressure. A qualitative causal model in the form of a concept map (Figure 3.6) is used as the basis for teaching the students the causal relationships between the different features of the system. For students who have problems with this high level view there are two more detailed maps which include more concepts and which also show the ways in which the different components affect one another (for example, neural or physical/chemical interactions). In functional/structural simulation, there is an attempt to explain the behaviour of the system in terms of its composition and the interaction of components or processes in a very general way. There are several techniques for achieving this and the general approach is called qualitative reasoning . The term, as Clancey observes, is a misleading one but has been adopted by the developers of functional/structural models and so will be used here. The technique has been used for investigating phenomena, for developing expert systems, for system design and various other purposes (Weld and de K.leer, 1 990) but many of the basic ideas were developed during research into computer teaching systems, particularly the work of Stevens and his colleagues on WHY, and the BBN team involved in the SOPHIE project.

BV BR-CNS CBV cc CO HR MAP PIT RAP RV sv TPR Blood Volume

B rain-Central Nervous System Cerebral Blood Volume Cardiac Contractibility Cardiac Output

Heart Rate

Mean Arterial Pressure Pituitary Gland Right Atrial Pressure Right Ventricle

Stroke Volume

Total Peripheral Resistance

+

Figure 3.6 Concept map of baroreceptor reflex (Khuwaja et al. , 1992, p218) The WHY system for teaching the basics of meteorology has been mentioned in Chapter 2. It has a script-like structure for representing causal and temporal relations and takes what Clancey would regard as a classificatory approach to processes. As Stevens, Collins and Goldin ( 1 982) observe, this produces a useful but limited model. By executing the script the system can demonstrate what is happening and can match the student's explanation. If the student goes wrong at some point or misses a step then the system can detect this and explain what should be done. However, it cannot deal with certain kinds of misconceptions and has difficulty explaining the inter­ relationships of the components of systems other than in a general fashion. For these reasons, Stevens and his colleagues propose another (functional) viewpoint for filling in these gaps in the system's explanatory and diagnostic capabilities.

In this alternative model, each of the elements is considered as a separate entity and the role that each plays in the process is described. In the rainfall domain the process of

evaporation can be viewed from this perspective. The ocean can be regarded as an actor with the role of providing moisture. Each of the actors has attributes that may affect the process (for example, temperature and altitude of air-mass). Functional relationships between factors explain the effects of interactions. For instance, the temperature of an air-mass is inversely related to altitude and so the air cools as it rises. The domain of electronic circuits is very different from that of meteorology, yet the workers on the SOPHIE project (Brown et al. , 1 982) came to a similar conclusion to the WHY researchers; that is, in order to provide insight into the operation of a model, the functioning of the constituents needs to be considered. The original SOPHIE was built around a quantitative simulation of electronic circuits. A knowledge base of information about the simulator was built on top of this that allowed the user to interact with the program and learn about fault detection in electronics. With this interpreter between the user and simulator the program could monitor the student's attempts at troubleshooting to see if slhe was using a logical approach, answer hypothetical (what­ if) questions about the circuit, and suggest hypotheses for the student to try.

As with WHY, the weak point of S OPHIE I was explanation. As Wenger notes in his assessment of the package: 'it makes no use of the kind of causal reasoning performed by human troubleshooters . . . causality is pedagogically important because it is the main ingredient of the kinds of explanation human students can understand' (Wenger, 1 987, p62). SOPHIE II (Brown et al. , 1 982) attempts to remedy this defect by providing a troubleshooting expert that can find faults in the circuit. Thus, for example, the user can introduce a fault and see how the program detects it. The procedure that the program follows, however, is quite brittle (essentially a decision tree). Although the system can provide causal explanations of its steps, these are pre­ stored and of a general nature and do not relate to the specific circuit. Also, the program cannot monitor a student's troubleshooting procedure if this deviates at all from its own actions and ordering.

In summary, SOPHIE 11 papers over the cracks in S OPHIE I. At the core of the problems with both systems is the use of a numerical program for simulating the behaviour of the circuit. In circuit analysis, as in many other domains, problem solving involves not just learning particular procedures but having a deep understanding of the system under consideration. The numerical simulator is essentially a black box and the interactions between components and the overall structure of the circuit has been coded in a mathematical form to which the user cannot relate.

S OPHIE Ill (Brown et al. , 1 982) provides an alternative approach. The simulator is replaced by an electronics expert that is knowledge-based and has a qualitative component. The expert has knowledge about the principles of electronics and uses this in conjunction with an automated troubleshooting expert to generate a situation­ specific model of a given circuit which is represented as a behaviour tree. In this tree

\

the module behaviours, structure relations and interactions between modules are stored. The information in this tree can be interpreted to provide a kind of simulation of the behaviour of the circuit.

The S OPHIE Ill approach still involved interposing a layer of processing between the description of the behaviour of the circuit and the user. Two of the researchers on this project, de Kleer and Brown, wanted to produce a more direct causal model of circuits and of other mechanistic devices, based on people's perceptions of how they operate. In a 1 9 8 1 paper (de Kleer and Brown, 1 98 1 ) they expound their views on how this might be achieved. They believe that the key to providing a model that people can relate to is to make it as close as possible to the mental model of the device that they have or that an expert might employ. It is claimed that this model is essentially qualitative and based on cause-effect principles.

When seeking to produce mental models, de Kleer and Brown aim for an ideal rather than the reality. As Norman ( 1 983) observes, most people's perceptions of how devices work are incomplete, unstable, confused, illogical and limited. de Kleer and Brown are looking for models that are descriptive, have simple to understand cause­ effect sequences of operations, and that correspond to the structure of the original in some direct fashion. The models must also be logical, general, robust and consistent. In order to preserve the structure of the original device each of the components of the referent must be modelled. The behaviour of each component must also be described, and, since each component may be a module in itself, this is recursive. A key concept is n o -function-in-structure. That is, when describing how a component works this description must not depend upon the behaviour of the system as a whole but only on how it is directly affected by� and affects other components of the system. As well as allowing localized behaviour to be explicitly described, the no-function-in-structure principle makes the system robust. If the circuit is altered or faults are introduced at a local level then only the portion of the model corresponding to the modified part of the original needs to be altered. Essentially, this is an object-oriented approach.

Related to the principle of no-function-in-structure is that of 'weak causality' which states that every event must have a direct cause. Again, the implication is that the

reasoning used fo\ determining the next state change should be localized and should not use any indirect information. A development strategy using these principles can, according to de Kleer and Brown, provide a qualitative model that corresponds closely to that an expert might employ and which could explain many aspects of the behaviour of a system.

To illustrate their ideas, they develop a sequence of models of an electro-mechanical buzzer which can be used to answer useful questions about its behaviour. It is important not only to be able to describe how it operates normally but to be able to determine what would happen if various changes were made such as reversing the leads of the battery or shorting the switch contact. Figure 3 .7, taken from Wenger ( 1 987), illustrates one model. Figure 3.7a shows the original circuit in typical schematic form, indicating the individual components and the electrical connections between them. Figure 3 .7b shows state diagrams for each of the components and for the three c omponents linked together. Lastly, in Figure 3 .7c a causal model shows how behaviour is propagated through the system as changes are made. de K.leer and Brown believe that these kinds of models, by illustrating how individual components interact and by making any implicit assumptions explicit, provide a useful target for students learning about devices.

There are two other main strands to qualitative reasoning that should be mentioned although they have had less influence on the research conducted in this thesis than has the work of de K.leer and Brown. Qualitative process theory (Forbus, 1 9 84) takes a different approach to conceptualizing mental models. Forbus focusses on processes, how they are enabled, relationships that are satisfied during their occurrence, and how they influence individual entities involved. In line with his work on STEAMER, Forbus attempts to describe systems at a conceptual level. Also he looks at systems as perceived by people in general rather than experts. This work is related to Hayes ( 1 985) research on naive physics.

Kuipers ( 1 986) takes a third view by looking at qualitative measurements within a system. The term qualitative measurement may seem like an oxymoron but refers to attempts to describe values of variables in a general way. For example, we might distinguish between an electrical current value that was negative, zero or positive without determining its exact value. Again, causal methods are used to show how the system changes, although, in this case, the causal reasoning is a form of constraint propagation. Values are propagated through the system and constraints determine which are "feasible. So the emphasis is on a mathematical model of the system rather than on the interaction of components or processes.

a) device

b) Device topology and component models (F1 =magnetic field 1 1 , 12, 13=electrical connections)

CLAPPER:

BATTERY:

c) causal model

CLAPPER

1 1 =0 or 13=0

11 and 13 are identical

Figure 3.7 Qualitative description of buzzer (Wenger, 1987, p72) based on de Kleer and Brown (1983)