In the present chapter, we shall describe the production system, which we introduced briefly in chapter 3, as a way of modelling human rule-based problem solving. In particular, we shall look at how humans, and machines, can call on internal representations of task domains, in the form of production rules, to guide them towards the solution to a problem, and on how they will act, at any step in a task, in specific ways in response to specific conditions.
By way of example, consider how you might go about diagnosing a fault in your hi-fi: you want to play a record, and you know that when you switch the power on, a little red light appears on the amplifier; if the light is on and the tone-arm on the record player does not respond to your pressing the button, you may know that it is a good idea to check whether the lead between the record player and the amplifier is properly connected; but on looking closely you notice that, in fact, the red light has not come on, in which case it is a sensible idea to verify that the hi-fi is plugged in; if it is not, you will do so, and try again; but if the hi-fi still does not work, you know enough about electrical appliances and domestic electricity to suspect that a fuse may have blown; and so on. At each step, you are invoking knowledge specific to the domain, checking whether a certain condition holds and, if so, performing the appropriate action. This kind of approach towards problem solving fits in well with a view of humans as information processing systems, which we explain more fully below, and contrasts with the state-space search procedures described in chapter 4.
At the end of chapter 4 we pointed out that state-space analysis may not be the most appropriate way of solving, or describing, some kinds of problems. An alternative is to look at why people behave as they do, and to describe the mental states and cognitive processes of people engaged in problem-solving tasks. It is impossible to observe those mental states directly; indeed, as we mentioned in the introduction, behaviourism, for very many years the dominant movement in psychology, went so far as to deny that there were any such things as mental states, and proscribed all talk that seemed to refer to `mentalistic' phenomena. But by the end of the 1950s there was active debate in psychology that ranged the behaviourists against a dissenting new generation of cognitivists, who maintained that it is possible to look into the `black box' of the human mind, and that psychologists should be concerned with a person's internal representation of the world.[+]
The cognitivist view was not new, in spirit at least: Gestalt psychologists had earlier in the century sought to explain problem-solving behaviour in humans and animals by positing the existence of certain kinds of internal representation (`Gestalten'). Two factors that weakened their position were the lack of, on the one hand, a systematic technique for revealing the details of subjects' internal representations of the world and, on the other hand, an explanation for the relationship between cognition and action: how actions might be controlled by an organism's internalized image of its world. Renewed interest in cognitivism came at the end of the 1950s with the publication of papers by Newell, Shaw, and Simon (1958, 1963a), and of a seminal study by Miller, Galanter, and Pribram (1960). With them came the notions of protocol analysis and production systems, which together would provide mechanisms for eliciting subjects' internal representations and explanations of how these are linked to actual behaviour.
Miller and his co-researchers looked to the work of ethologists on instinctual behaviour in animals: an animal, it was argued, `tests' to find whether a certain environmental condition holds and, if so, it performs the appropriate action. Different external conditions will trigger different associated responses of the form
if condition 1 then action 1if condition 2 then action 2
if condition 3 then action 3
if condition 4 then action 4
Under this view of behaviour, there is no longer a problem of how cognition is related to behaviour: conditions and actions are paired together in such a way that, if the former is satisfied, the latter is automatically executed.
But humans do not just respond to the environment; they also devise plans for negotiating their way through the routine problems of everyday life. A plan is a program for behaviour that is triggered when certain conditions in the external environment are satisfied. For example, I have a plan for doing my shopping on a Saturday morning: picking up my cheque book, and walking to the shops if the weather is fine, taking the car if not. If walking, then I will need a sturdy shopping bag; otherwise I can just pack my shopping into the luggage compartment of the car. If driving, I will need to stop on the way at a service station if I need gas; go to the open market for fruit and vegetables, and the supermarket for pre-packaged goods; pack the bags at the supermarket checkout in such a way that the larger and heavier items are at the bottom, the smaller, lighter items at the top; and pay the cashier. Back home, I put the frozen items in the freezer, the perishable items in the refrigerator, the vegetables in the vegetable rack, and the rest in cupboards. But I do this with very little reflection; I do not need to re-plan my morning's activities each Saturday, since I have done it all so many times before. Thus I have a standard plan, or perhaps an interlinked group of plans, which `as a rule' determine my behaviour. Such plans can be represented as condition-action rules, as above, or schematically as AND/OR trees and graphs. An AND/OR tree (or graph) is a way of representing solutions to problems by decomposing them into subsets of increasingly smaller problems. Arcs joined by curved ties link together AND nodes, and indicate that, in order to accomplish the goal in the parent node, each of the subgoals must be accomplished; arcs not so joined point to OR nodes, indicating that the goal in the parent node can be accompished by accomplishing either of the subgoals. The difference between a tree and a graph is that each node of a tree has at most one parent node, while a node of a graph can have more than one parent. Figure 7.1 represents a fragment of the AND/OR tree that maps out my plan for Saturday morning shopping. A translation of the tree into English prose might begin along the following lines: ``In order to go to do my shopping, I must either drive to the shops, OR walk AND take a strong shopping bag ...''
[IMAGE ]
Figure 7.1: An AND/OR tree for a shopping plan.
Complementing the contribution by psychologists, the computational metaphor of the human mind as a symbol-manipulating, information-processing system, pioneered by Newell and Simon, provided a way of talking about elusive internal cognitive structures and mechanisms and hence about the structure of human behaviour rather than about the structure of the problem alone. What was still needed was some method of `observing' human thought processes. Since they are not directly open to inspection, however, a less direct but still empirical method was required. One way, which we briefly mentioned in the last chapter, is to measure subjects' reaction time to stimuli. But it is also possible to study mental events by having subjects report what is going on in their minds during the performance of the task itself.
![[IMAGE ]](img3.gif){kind=link}