4.3 Computational Models in Artificial Intelligence
4.3.1 Classical AI
In the early days of AI, Allan Newell and Herbert Simon pioneered an approach where protocol analyses (verbal reports of thought processes) of human problem-solving were used as the basis for designing AI systems like their General Problem Solver (GPS) program (Newell and Simon 1961). This approach came to be known as symbolic or classical AI. Classical AI uses logical rules to manipulate syntactic structures, and hypothesizes that a cognitive system is like a programmed computer (or according to some is a computer). This approach has included many practitioners working on a variety of problems, but I’ll take Newell and Simon’s work as a representative early example.
Newell and Simon postulated that cognitive behavior is produced by elementary infor- mation processing over symbols, and that neurophysiological mechanisms in turn produce these information processes. This is represented in Figure4.1. GPS is assumed to share with human cognition the level of elementary information processing, intermediate between its program output and hardware instantiation. The claim that any intelligent system shares this intermediate level is codified in their physical symbol system hypothesis, which states that, “A physical symbol system has the necessary and sufficient means for general intelli- gent action” (Newell and Simon 1976). The defense of this postulate “lies in its power to explain the behavior” (Newell and Simon 1961), they claim, and early successes with their method led them to conclude that GPS could be considered a theory of problem-solving behavior.
Note that in this approach, theory development follows simulation. In physics and eco- nomics, theory usually comes before simulation, and forms the basis for it. In Newell and Simon’s case, a set of fundamental assumptions did go into the construction of the simula- tion, but the point of simulating was to discover what makes problem solvers tick. Their method is to start with an educated guess about underlying structure, see what their guess
Figure 4.1: Levels in an information processing theory of human thinking. Reprinted from Newell and Simon(1961). Copyright c 1961 American Association for the Advancement of Science, used with permission.
generates when simulated, alter the model based on this feedback, and repeat.
Newell and Simon make several interesting methodological remarks. They cash out what they mean by a theory in the language used to describe theories in physics: “From a formal standpoint, a computer program used as a theory has the same epistemological status as a set of differential equations or difference equations used as a theory”(Newell and Simon 1961). They then go on to compare boundary conditions governing the applicability of differential equations in physics with environmental inputs determining the successive states of an AI program. What they mean by theory then, is the Syntactic View.3 Newell and
Simon acknowledge that a program being a theory, in this sense, does not necessarily make it a good theory: “How highly we will prize this theory depends, as with all theories, on its
3 Since every program computes some computable function, a program like GPS certainly could be
represented in the form of law-like statements. Sun(2009) goes to the trouble of spelling out some of the details of how to do so.
generality and its parsimony” (Newell and Simon 1961). Beyond these criteria they are, of course, concerned with whether the same elementary information processes show up in their participants’ protocol reports as in their program traces. More detailed criteria like realistic reaction times were not a concern, at least at this early stage.
In addition, Newell and Simon acknowledge the role of neurophysiology in a complete theory of cognition; this is significant, because later defenders of classical AI reject the idea that neurophysiology plays any role in the study of cognition. Newell and Simon’s work on GPS is only concerned with discovering what the elementary information processes are, based on protocol reports, and checking whether these processes can generate convincingly similar output when programmed into a computer. Nevertheless, they note the need for “a second body of theory ... to explain information processes on the basis of neurological mechanisms” (Newell and Simon 1961). They justify setting this work aside on methodological grounds. They suggest that, “Tunneling through our mountain of ignorance from both sides will prove simpler... than trying to penetrate the entire distance from one side only” (Newell and Simon 1961), presumably meaning that their psychological study could be done top down, while the physiological study would be done bottom up.
Their belief that internal psychological processes consist of symbol manipulations also plays a role in this methodological choice. If the physical symbol system hypothesis is true, then there shouldn’t be any problems making the top-down and bottom-up analyses meet in the middle. Without this hypothesis, supposing that psychological systems must be supported by symbolic processes like those used in programs like GPS based on their effects being similar, would be a shaky inference from like effects to like causes. It is because of this assumption, based on introspection and common-sense psychology, that their methodology seemed justified. Without the assumption of a shared intermediate level, it would be hard to justify an exclusively top-down methodology.
Although the physical symbol system hypothesis was originally presented as an empirical claim to be tested, and the choice to work top down from behavior to information processes began as a proposal for dividing labor, these aspects of the classical AI approach became entrenched by the 1980s. A number of developments played a role in this entrenchment, including: Minsky & Papert’s (1969) effective, if unfair (see Boden (2006)), quashing of
Rosenblatt’s (1958) connectionist approach, plus Fodor’s Language of Thought hypothesis (Fodor 1975), and his arguments for the autonomy of psychology based on multiple realiz- ability (Fodor 1974).
In the meantime, theory had taken on a much more nebular status in philosophy of science. Particular AI programs designed to perform specific tasks were (and are) still referred to as theories, but this no longer harkened back to the Syntactic View, perhaps because this view was no longer so well received among philosophers of science. More often, by the 1980s, collections of basic assumptions or conceptual frameworks were referred to as theories, as in the computational theory of mind.
The classical AI approach has proven extremely useful for modeling some examples of intelligent behavior, like playing chess and solving logic problems. But symbolic AI didn’t pick up again the other half of the problem: figuring out how physiology gives rise to elemen- tary cognitive processing. This neglect makes it particularly unsuited for figuring out how cognitive phenomena are achieved by the brain, which is the goal of cognitive neuroscience.