The guiding idea behind much of the research and theorizing discussed in earlier lessons can be stated simply :
The human cognition system is best viewed as information -processing system. This idea, which constitutes the foundation for cognitive science, has been developed most fully in the field of computer simulations. The chief aim of this area of research is to construct computer systems that mimic or simulate both overt human behavior and the cognitive process that underlie it. An adequate simulation of problem-solving, for example, would not only solve all of the problems that people solve but would also solve the problems using the same kinds of representations and strategies that people use.
Cognitive scientists try to devise computer simulations for several reasons:
~ First, they want to test the sufficiency of their theories. As mentioned in the 'discussion of global, oppositional models of cognition, a theory that is embodied in a computer program can be tested by seeing whether the theory can do what it is intended to do, for example, read text and answer questions covertly.
~ Second, theorists want to define terms such as understanding and representation as explicitly as possible. Theoretical explicitness is no small advantage, for theories are testable only if they are explicit.
~ Third, and most important, a computer simulation is a genuine psychological theory that helps to guide research and unearth useful concepts.
Computers and people obviously differ in their physical composition yet they both manipulate symbols, and they may perform a particular act by following logically similar steps and procedures. So to the extent that a simulation models human performance accurately, then to that extent the simulation is an adequate psychological theory. In devising computer models, theorists often break new ground, using information from other disciplines such as computer sciences. If, for example, cognitive psychologists have not clarified the types of representations people use, then they may turn to computer scientists to pick up ideas about representations. In practice, these ideas often turn out to he useful in psychological theories. Indeed, psychological theories, emphasizing, schemata and prepositional representations owe a large debt to research in computer science. Because successful computer simulations do introduce powerful new concepts, they are genuine theoretical advances, not just noteworthy technological achievements. Let us briefly examine one of the most prominent efforts of simulation.
THE GENERAL PROBLEM SOLVER (GPS)
The GPS was designed to simulate human problem solving in a broad range of tasks (Newell and Simon, 1972). It includes a long-term memory that stores semantic and world knowledge as will as operations for using that knowledge. GPS also includes a short-term memory (STM) of limited capacity that necessitates serial rather than parallel processing. The STM is a workspace in which objects are compared and decisions are made about how to act next. The GPS system works by examining a problem space. The problem space is the internal representation of the problem, and it consists of all of the legal states that the system could enter while working on a problem. Using the heuristic of means-end analysis, the GPS compares the initial state in a problem to the goal state. Then it selects a move that will reduce the difference between the current state and the goal state. , Most of the knowledge included in GPS is in the form of production rules, which instruct the system to take a particular action if a particular condition has been satisfied (Newell, 1973; Simon, 197~). Here are some examples of production rules:
~ If the light is red, then stop
~ If hungry, then eat some food
~ If tired, then rest
Each rule consists of a situation-action pair, and each rule incorporates a tiny bit of knowledge about how to act under particular conditions, either inside of or outside of the system. Of course, a powerful system must have many thousands of production rules.
Researchers must have a good idea of the steps that people go through in solving problems, in order to evaluate whether GPS solves problems in the same way that humans do. For this reason, investigators use the method of protocol analysis, in which subjects solve a problem while thinking aloud. The verbal protocols of the subjects often reveal the problem space the solver is using, the state of knowledge the solver is currently in, and the operations the solver in using to advance toward the goal state. Having collected this information the investigator can then evaluate the degree is similarity between the problem spaces, the knowledge status, and the operations used by humans and those used by GPS. This evaluation is difficult to make since there are no accepted rules for deciding the degree of similarity. And the intuition of the investigator plays a key roe in all stages of the method of protocol analysis. Nevertheless, the method has proven to he useful.
Computer Simulation in Well-Defined Problems
In several will-defined problems researchers develop computer simulations. The researcher's task is to create computer program that can solve problems. According to this, by developing the instructions a computer must execute to solve problems, the researcher may better understand how humans solve similar kinds of problems. Based on early work in computer-simulated problem solving Allen Newell and Herbert Simon ( 1972) developed a model of problem solving.
According to the Newell-Simon model, the problem solver must view the initial (problem) state and the goal (solution) state within a problem space-the universe of all possible actions that can be applied to solving a problem, given any constraints that apply to the solution of the problem. According to this model, the fundamental strategy for solving problems is to decompose the problem task into series of steps, which will eventually lead to the solution of the problem at hand. Each step involves a set of rules for procedures (operations) that can be implemented. The set of rules in organized hierarchically into programs containing various internal levels of sub programs (called "routines" and 'subordinates"). If a computer is provided with a well-defined problem and an appropriate hierarchy (program) of operations organized into produced algorithms, the computer can readily calculate all possible operations and combinations of operations within the problem space and determine the best possible sequence of steps to take to solve the problem.
Although GPS uses a number of difficult methods to solve problems, these methods . generally draw on a single heuristic for problem -solving. The heuristic GPS uses is means-ends analysis, which involves solving problems by successively· reducing the difference between the present status (where you are now) and the goal status (where you want to be). Fig. 3l.a shows a schematic flow chart (a model path for reaching a goal or sol,ling a problem) for how GPS can transform one object (or one problem state) into another using means-ends analysis.
In several well-defined tasks, such as cryparithmatic problems and the hobbits and-ores problems, the performance of the GPS system resembles that of adults (Atwood and Poulson, 1976; Newell et al, 1972). The performance of GPS deviates what from that of people in being too single minded, processing in a strictly serial manner, and relying more heavily on means-end analysis than people do (Greeno, 1974). But its overall achievements are noteworthy, particularly because it performs such diverse tasks using a small under of general elementary processes such as comparing items in short -term memory. Unfortunately we do not yet know whether the system can be extended in ways that allow it to simulate human performance in more complex tasks, for example, tasks that call for intuition and creativity. At the very least, GPS must be supplemented by programs that form representations for problems, since GPS lacks this ability (Hayes
and Simon, 1976).
For our purposes, however, the potential limitations of the GPS system are of less interest than are those of the computer simulation approach in general. Let us look into the limitations in order to identify significant challenges to the entire information-processing outlook.
THE LIMITS OF SIMULATION
Attempts to simulate cognition via computer face serious methodological and theoretical obstacles. A severe methodological problem, stemming from the privacy of cognitive processes, is that it is unclear how to assess whether the performance of a computer simulates in .detail the performance of people. Protocol analysis can reveal some of the conscious steps that people take in solving problems (Ericsson and Simon, 1980). Using protocol analysis, we can assess the similarity between the steps taken by a computer and the conscious steps taken by people. But non conscious processing may contribute to human problem solving, and the conscious thoughts that people have may be products of underlying processing that lies outside the scope of protocol data with data concerning the performance of a computer cannot possibly tell us whether people and computers used similar underlying processes in performing a particular task. Even if we knew that a person and computer had gone through the same global, macroscopic steps using similar underlying processes. Until this problem has been resolved we cannot determine exactly how well computer performance simulates human performance.
A second problem, related to the first one is that human cognition is guided by an immense amount of background knowledge. This knowledge lies on the fringes of consciousness and is applied before logical, conscious planning begins. Further, this knowledge does not enter directly into the representation, yet it shapes the properties of the representation. By analogy of the figure-ground relationship in perception. the background knowledge & the ground in which the representation is formed. Since this background knowledge guides the formation of representations, and since the form of the representations influences problem solving, a comprehensive theory must specify what the background knowledge is and how people use it. Computers have been programmed to simulate some aspects of human learning (Anderson, Kline and Beasley), 1979~ Feigenbaum, 1963). But research into learning is in its infancy, and it remains to be determined whether computers can learn background knowledge in the same ways in which humans do.
Another problem is that computers programs tend to become so complex that they are difficult to understand (Weizenbum, 1976). The best of programmers is rarely so lucky as to have a carefully written program work correctly on its first run. Usually, programmers discover flags or "bugs" and while debugging an intricate program they introduce large number of steps into the system. Ironically, the program may work and the programmers may not understand why it works. And when a program is loaded with minor details, it becomes difficult to discern the main theoretical assertions of the system (cf. Smith, 1978). Finally, let us consider the last problem, that computers lack emotions. As the cognitive and affective processes are assumed to be independent, the absence of emotions may .not really pose problems. But even if they are independent they are said to l influence each other extensively (Bower, 1981 ). Large evidence indicates that human decisions depend upon both emotional and cognitive factors and that the emotional factors operate more quickly than the cognitive factors (Zajonc, 1980). Because the emotional factors operate first, they can influence, perhaps even guide, cognitive processing. Computer systems cannot model such a scenario because they have no emotions. Further an emotion cannot be reduced to a symbol structure of the sort that guides activities of a computer. Emotions have a bodily component that computers simply lack. For this reason, computers would not understand romantic poetry and art works in the same way as the human do. This brings out a significant challenge for the future, to construct computers having bodies and to determine whether the interplay between human emotion and cognition can be simulated by a machine.
From the above comments, although the information-processing outlook has been useful, it reminds us that it may not be able to accommodate all of the important facets of human cognition. People may always have abilities that outstrip those of machines, so it is important not to cast human cognition entirely in the image of the computer. But by the same token, cognitive psychology has advanced by uniting ideas from diverse disciples, and we may expect the field to advance further by building upon the insights from computer science.
SUMMARY
Cognitive psychology has drawn some of its most useful theoretical constructs from theories attempting to simulate human cognition via computer. So far, investigators have constructed computer systems that perform like people do in a moderate range of problem-solving tasks and learning tasks. These successes speak well for the view that the human cognitive system is a type of information processing system. This view has proven to be useful in cognitive psychology and will no doubt continue to be so. Yet much more research needs to be directed at whether computer modes can simulate the background knowledge that people have or the affective -cognitive interactions that pervade our lives.