Chapter 6 Knowledge-based Simulation for Teaching
6.4 Modelling behaviour
6.4.2 A behaviour model for teaching
In the simulation of a physical system the general flow, at the most simplistic level, can be considered to be: situation --7 action --7 situation --7 action . . . . At each stage, a new
situation enables an action to be executed and each action causes a change in the situation. There are some similarities between the causal chains that are modelled in physical systems and the way that humans act and respond. How far this analogy can be taken remains an open question, both for philosophers and psychologists. A philosophical perspective is given by Davis ( 1979) who believes that causal events have an inevitability and are logically deducible, whereas human actions have some element of choice. He concedes, however, that although human behaviour is not totally predictable, it can be surmised and can usually be explained once it has occurred.
When considering human behaviour, the psychologist John McClure ( 1 99 1 ) distinguishes three aspects: intentions, reasons and causes. Intentions, he states, 'articulate a plan of action' and may be conceived as the class of response one might give to a question such as 'what are you planning to do?'. Reasons give a motive or rationale for an action and may be typified by answers to the question 'why are you doing this?'. Lastly, causes 'comprise a mechanical explanation of behaviour' and correspond to replies to the query 'what makes you do this?'. This extends the earlier ideas of, for example, Davis ( 1 979) who distinguishes intention and reasoning. He characterises intention as how the agent thinks of the action, relating it to the notion of purpose
(although the latter term is sometimes associated with the idea of reasons). Davis uses the word purpose exclusively to mean 'what the agent aimed for'. Reasons are more related to beliefs and desires. More precisely, beliefs are held that a particular desire can be fulfilled by a particular action.
Some of these ideas are incorporated into the model of behaviour considered here. In scenarios where actions are being executed by humans, it is useful to distinguish the mental states of these actors. It is assumed that the way that someone feels will influence how they behave. This augmented model of cause and effect is shown in Figure 6.7.
outside
Figure 6.7 Interaction of situations, mental states and actions
The shaded area shows the scope of simulation of interpersonal relationships. Outside events occur and have to be considered but their justification cannot necessarily be included in any detailed fashion. To be able to include a full qualitative simulation would require an intricate model of the whole world on the scale of the CYC project (Lenat and Guha, 1 990). However, the reaction of people in the system to situations influenced by outside circumstances can, arguably, be simulated. The initial impact is upon their mental state, and this then will result in some action that changes the current situation. The kinds of action that may be taken include verbal as well as physical.
Even within this simple framework some feedback can be given. The different kinds of guidance are summarised in Figure 6.8.
operation possible feedback
actions resulting in state change alternative results of action indicated
desirable action recommended
action given to achieve specific state
states initiating mental states effect of current situation on mental state of actor current mental state of actor
situation that might induce given mental state
mental states initiate actions what person is likely to do
It might appear that the model given in Figure 6.7 takes a behaviourist approach where a situation causes someone to react in a certain way that then changes the situation. This is not, in fact, the case. Rather than thinking of the diagram as denoting sequencing of operations, it is perhaps best to think of it as a petri-net (Reisig, 1985) where branches show direct influences but where a specific sequencing is not implied. For example, although a situation may change the mental state of a person, this does not necessarily lead immediately to an action. A person's mental state may incorporate short (or even long) term goals that dictate their actions over a period of time. The changing situation may give them feedback on whether their aims are being achieved.
Another assumption implicit in this model is that a person's actions depend not upon the situation but upon their perception of the situation. This will be interpreted using their current beliefs and predilections to determine their reaction. These distinctions may be crucial to help explain seemingly irrational behaviour.
More information is needed in order to animate this model. In particular, the mental states of actors within the system have to be described in enough detail to be able to determine how their behaviour is affected. A detailed psychological model of behaviour of human beings is, of course, currently out of the question. It is argued here, however, that known patterns of behaviour can be simulated and corresponding plausible development of situations can be presented. The aim is not to predict how everyone will behave in every circumstance, but to be able to string together credible sequences of events based on knowledge of human behaviour. In addition to being credible, it is hoped that the knowledge that has been gleaned by psychologists, social scientists and so on can be conveyed at some level and in sufficient detail to enable users to gain a better understanding of certain types of scenarios and, perhaps, to enable them to consider alternative courses of action.
To be able to develop a working model, various simplifying assumptions have to be made. In this process, as in any modelling exercise, the aim is to ensure that the baby does not get thrown out with the bath water. The main model used as a basis is the one developed by Schank and Abelson, described in Section 6.4. 1 .
First it is assumed that a profile can be generated for each of the actors within a system. This will include information about relevant beliefs, feelings, and other aspects of character. From this information potential goals can be hypothesized. To achieve these goals specific plans are made. Lastly, the plans involve carrying out sequences of actions. Things do not always run -so smoothly, of course. For example, other people's goals may clash with our own or the plans that we have may need to be changed due to
unforeseen circumstances. Although every eventuality cannot be covered this gives, at least, a framework for human behaviour. Now, consider each of these ideas in more detail.
Much of what is required in a person's profile is incorporated into Schank and Abelson's life themes. People's behaviour is influenced by their personal qualities, their ambitions, .
their life styles, what approval they seek and by their need for physical sensations. To this may be added their belief system. According to Bootzin and his co-authors a person's beliefs consist of 'knowledge structures about objects and events' (Bootzin et al. , 1 986, p607). Along with these they include motivations which they define as the reasons for beliefs. Clearly, beliefs will be personal and may be totally at odds with those of the general population. Motivations, therefore, provide personal justification at some level.
Bootzin and his colleagues note that one cannot necessarily predict behaviour from beliefs since, for example, people may be perverse or there may be some unknown interplay between what they believe and what they want to achieve. However, for the most part, a good idea of how they are liable to behave can be predicted from their beliefs.
Assuming that a person's profile can be used to determine what they would like to do and be, goals can be generated for people, both long term and short term. These may change due to circumstances (for example, if an actor believes a goal is no longer realistic) or if the attitude of the actor changes. Goals may be unattainable, fanciful, limited or destructive depending upon the character of the actor.
Besides the individual goals of an actor, Schank and Abelson's expectancy rules give some guidelines for what they might try to achieve: for example, to show gratitude or to exact revenge. Expectancy rules provide the glue for fixing together the whole system. Goals may be split up into sub-goals, or instrumental goals for achieving the main one. For example, seeking revenge may involve the instrumental goal of bankrupting the unfortunate target of one's wrath.
Also, each actor may have different views on the best way of attaining their goals. The variation in methods of achieving goals may be due to different beliefs about the current situation, different capabilities of the actors, different moral codes, etc. Given a specific goal for a particular person in given circumstances it is possible to formulate plausible means by which the person may seek to achieve the goal. This leads to the idea of a plan.
A plan may be good, poor, watertight, flawed or just plain stupid. However, it is a means by which the formulator intends to aim for a particular end. It may deal with all
kinds of eventualities with provision for modifications as circumstances change. On the
other hand, it may fall down at the first hurdle and be abandoned. In any case it provides
the guidelines for the actions that a person carries out. This corresponds to McClure's notion of an intention.
At the bottom level of this hierarchy is the action. An action is motivated by some aim. It may almost be a reflex response based on immediate circumstances or it may be part of a thoughtful strategy for achieving a particular goal. In Figure 6.7 the action of an actor was depicted as being triggered by their mental state. This mental state is presumed to incorporate long term as well as short term goals and plans.
At a high level, a scheme that culminates in action is needed. Such a scheme is illustrated in Figure 6.9. Profiles are assumed to consist of themes and beliefs. The types of goals, taken from Schank and Abelson's book (Schank and Abelson, 1977) are self-explanatory except for instrumental goals (intermediate goals) and delta goals (changing the current state). As with Figure 6.7 this diagram shows the general impact of one stage upon another and does not imply an absolute sequencing in particular circumstances. Obviously, the current situation may change what a person does, but each level is more resistant to change than the one to its right. Plans can be changed according to what happens, goals are less likely to be influenced by events and a person's profile is unlikely to be affected except in unusual circumstances ('seeing the light', disillusionment, etc).
Life themes Role themes Interpersonal themes Beliefs Satisfaction Enj oyment Achievement Preservation Crisis Instrumental Delta Named plan Novel plan Script
Figure 6.9 Profiles, goals, plans and actions
The models presented in Figures 6.7 and 6.9 contribute to the development of a feedback system. Figure 6.7 indicates the various actions that are enabled, based on the current mental state of the actors. So there may be several plausible alternatives that a person
may choose from (that is, enabled actions within the current state of things). Thus any one of a number of actions may be taken. If the user assumes the role of an actor then a choice of the enabled actions may be presented. Any of these can be justified via the preconditions for a particular action. The effects can be shown via postconditions.
The higher level description in Figure 6.9 can be used to show longer term and more considered reasons for behaviour. If it is required to show why a person may have carried out a particular action rather than any other feasible ones then this can be justified by the plan that this person has. The plan can be shown to have emerged from a general goal, and this goal can be justified in terms of the actor's beliefs and attitudes. As well as justifying what actors have done, this model enables the computer to suggest actions that might be carried out which are consistent with the user's goal (or any otner goal suggested by the system), or an attempt can be made to infer information about users from their behaviour. Their plans might be deduced from their actions, their goals from their stated plans, and their attitudes from their stated goals. In this way, a meaningful dialogue can be built up and tailored to the needs of the user.
Not much has been said about specific circumstances but these obviously play some part in how someone behaves. To some extent a prediction of what a person will do can be based on the current situation (or more precisely upon their perception of the current situation) but this may be influenced by factors other than just their attitudes and beliefs. Stock situations can be represented by scripts and these can have a great deal of influence on how people react. In addition, the role that someone sees for themself (Schank and Abelson's role theme) can have a significant impact. In a realistic model these have to be included. Care is needed when employing scripts, however, since, to an extent, they gloss over the details of why actions are carried out. A successful Script Applier Mechanism (SAM) was developed by Cullingford ( 198 1 ) , but, as noted by Dyer ( 1983), SAM cannot answer questions such as why someone did something.
Comparing the above model with that developed in Chapter 4 it is clear that it is going to be more difficult to implement. Any given situation is going to be more complex to analyse because of the requirement to consider individual attitudes, beliefs, reasons for behaviour and so on. It is also clear, however, that there are similarities that can be exploited. The causal semantics involving behaviour can be likened to the cause-effect model developed previously. Each situation that is considered will contain the extra component of mental states of the actors but, again, plausible reasons can be found for the choices made. This extra information about people's plans, goals and expectations would hopefully provide feedback to guide the user. The plausible actions that the user might take (in line with their profile and goals) could be considered to be the domain
, ,
constraints. Specific recommended courses of action could be provided by an overlay on
this, and is analogous to the task specification developed in Chapter 4.
6.5 Summary
A scheme for modelling human behaviour and interaction within a framework of knowledge-based simulation has been developed. In such an ill-structured domain as simulation of human behaviour much simplification is required to enable any kind of computer model to be constructed. The aim has been to simplify the system without losing the main driving features of human interaction: attitudes, beliefs, goals and plans. A prototype system based on Lenat's scheme for event description has been presented. A more detailed model has also been described that is based on Schank and Abelson's research into human knowledge structures. It is proposed that such a model can be adapted for teaching purposes.
Using the attitudes, beliefs and goals of the actors to drive a system based on the above model has several advantages:
a high level of transparency is maintained so the user can see what is happening; since the plot line is not hardwired in, there is flexibility in what can occur;
the domain expert can be involved in the package development to a greater degree;
the methods of Chapter 4 for highlighting routes in the system can be employed,
enabling the teacher to focus on specific situations;