Artificial Intelligence and Economic Analysis.

The 9 papers collected in this volume were prepared for a conference on artificial intelligence (AI) held at Manchester Polytechnic in December 1990, whose purpose was to examine the potential relevance of this line of research for economics. The book purports to focus on 3 lines of AI orientated research with potential economic applications. That is, applications that attempt to undermine neoclassical economics, applications intended to reinforce it, and those that by-pass the theory by presenting what the authors call "new modelling techniques." In this, the editors deserve a cheer for trying, for I believe that AI research does hold great promises for the economic model of rational choice. Having said that, I hasten to add that, as most economists know from experience, big projects that purport to give us a more robust theory are always risky; and this one is no exception.

Since there are no clear demarcations, I presume that Part I comprises the Introduction and the next two chapters. In chapter 2, Robin Marris who (most economists would remember as the man who tried hard in his career to inject greater realism in economic and financial discourses) kicks off with a rather superficial discussion on the nature of intelligence, providing unfortunately no more than food for thought. In the next chapter, Moss focuses on AI models of complex (sic) economic systems by considering applications of two machine-learning procedures. In what I surmise as Part II, Rae and Reynolds present AI modelling techniques within a behavioral perspective. Romeo and Moss attempt, in chapter 5, to contrast the performance of expert-systems and standard economic theory in specific situations. In chapter 6, Hey and Reynolds examine, with two experiments, the extent to which individual behavior coincides with the predictions of neo-classical economic theory. In the next two, Paul Stoneman investigates, albeit in general terms, the relationship between AI and the economics of technological change, while Huw Dixon offers some thought on economic theory and AI. In the last part, Anderlini considers players' computational ability and reviews the literature on the theory of games where players are able to handle those tasks which are computable within Church's thesis, and finally; the last chapter considers potential contributions of AI to macro econometric modelling.

The reading is not always smooth due to the unevenness of the chapters as well as the amateurism of the presentation. But more so the potential contribution of the book, if any, is not obvious to this reviewer for a number of reasons. As everyone knows, to solve the economizing problem, economists have developed what is commonly known as constrained-optimization algorithms (COA), and everyone also knows that at times problem spaces are simply too large for the application of COA; hence, it may be said that the limit of the orthodox approach to rationality is well-known. Beyond bounded rationality then, remaining problem areas are: knowing whether or not aided-solutions are optimal, and where agents face problems that are not well-defined. But these are precisely the weaknesses of the AI techniques covered in the book. I guess that before economists could become convinced of the benefits of such techniques, advocates must demonstrate their pay-off in terms of these remaining problem areas. However, I am not myself convinced that the book does that. Moreover, despite many claims, the authors do not really deal squarely with the standard definition of complexity. As far as I am aware, the study of complexity concerns the search for the few simple principles that might govern the development of higher-order structures, i.e., an anti-second law of thermodynamics offensive, if you will. In this regard, the editors pay lip service to what they term environmental (see d) below) and outcome (say random determinism) complexities, and still manage to leave the reader eager to know exactly how AI techniques could be brought to bear on these. Indeed many of the examples are trivially simple, while having no relevance to how economic agents form their preferences or to the consistency of the decision-making process; two necessary clarifications before economists could venture beyond the axiomatic base of their rational choice model.

This is not to say that I condemn the effort. On the contrary, I believe that AI has great relevance to economics, but not necessarily within the limited framework dealt with in this book. I have the impression that by "complexity" the authors really have "difficult problem spaces" in mind. However, even there the application of the techniques they refer to is not so straightforward. To stress the point, consider a subset of difficult problem spaces:

a) determine the number of periodic solutions of a planar autonomous differential equation system in the form of |Mathematical Expression Omitted~; call this the Hilbert 16th problem;

b) compute the general equilibrium of the U.S. economy (a Radner |1~ problem);

c) keep track of the altitude, speed and direction of every aircraft in the vicinity of O'Hare Airport (a Simon |3; 4~ problem);

d) find the optimal first move in the set of all possible games of chess of 40 moves or less.

The reader will notice that in a) the agent really faces a multiplicity of solutions, in addition to limited computational capacity; b) and c) differ by the incompleteness of the information set in b); whereas in d), the problem is computable in principle but it is uneconomical to attempt to do so. The AI techniques the authors focus on apply only to the Simon problem c). Though c) is relevant, it is by no means more so than say, a) or b). The main drawback of the book to me stems from the authors' use of the term "expert-systems" interchangeably with AI. Expert-systems are intelligent computer programs that use knowledge and inference procedures to solve type c) problems, i.e., devices that enhance human computing capacity and expertise. They offer no help in types a), b) and d) problems. To be more explicit, consider a natural limit to "unlimited computing capacity" exemplified by problem d). It is estimated that there are 25 x |10.sup.115~ games in d), a number considerably larger than the number of electrons in the known universe (|10.sup.79~). It may then be said that even unlimited computing capacity would not do, for the fastest super computer we now have cannot solve problem d), not even in 5 billion years. Therefore, the authors should have been careful enough to emphasize that the present pay-off of expert-systems is that they may allow non-experts to arrive at what is now considered expert solutions, and that they cannot guarantee economic optimality nor can they enhance theoretical understanding. Thus while they may make a difference in decision-making in disequilibria and in strategic environments, great care must still be exercised so as not to swing economics toward OR.

It would have been more fruitful, in my view, to reserve the term AI for neural research and the construction of genetic algorithms and artificial life systems, where resides the hope of understanding preference formation. Many philosophers of science (see Rosenberg |2~) would agree that beside the problem areas alluded above, the great challenge to economic theory is to account for the impact of intentionality on rational decisions; that is the extent to which beliefs and desires influence economic decisions. I am inclined to believe that a greater understanding of intentionality and its relationship to economics will do more for the model of rational choice than the enhancement of agents' computing capacity. Furthermore, my belief is that this will not come about until we get a grasp on the functions of the four million modules or the |10.sup.10~ neurons of the association cortex of the brain, as now being actively pursued in cognitive neuroscience laboratories in France and Japan, or at the Santa Fe Institute in the U.S. In this regard, I dare say that in the end and in line with the above definition of complexity, it might eventually be found that the theory of sequential machines with integral transform encoding and decoding maps is the most promising avenue. Unfortunately, the book ignores these exciting developments. Let us hope that the authors are saving them for their future project.

References

1. Radner, Roy, "Competitive Equilibrium Under Uncertainty." Econometrica, 1966, 31-59.

2. Rosenberg, Alexander. Economics: Mathematical Politics or Science of Diminishing Returns. Chicago: University of Chicago Press, 1992.

3. Simon, Herbert A., "Rational Decision Making in Business Organizations." American Economic Review, 69, 1979, 493-513.

4. -----, "Decision Making and Problem Solving." Interfaces, 1987, 11-31.