Towards a not-too-rational macroeconomics.

• Author: Leijonhufvud, Axel

1. Introduction

   At UCLA we are establishing a Center for Computable Economics. I have been very much involved in this effort. This may surprise you. That a non-mathematical macro/monetary economist should become an enthusiast for this project is one of those things that "do not compute." But then, as I will explain, we take a special interest in things that do not compute.

   One may come to Computable Economics by many intellectual routes. I will trace my own, not because it makes a particularly edifying story, but because it will tell you what kind of problems I hope we can make progress on by developing the field of Computable Economics. But before we get to that, I had better explain what I mean by Computable Economics.

2. Computable Economics

   The computer is now being used in a wide variety of fields to model and to explore the properties of complex dynamic systems. We believe that this approach has a big potential payoff also in economics.

   In macroeconomics, to take an example close to my heart, the last ten or fifteen years have seen the almost total abandonment of static in favor of dynamic models. Dynamical systems, however, have to have a very simple structure if one is to obtain closed form solutions. The core of this modern macroliterature consists of representative agent (or social planner) models, where the motion of the entire system is given by the solution to a single optimization problem. It is possible to go a bit beyond the representative agent and introduce, for instance, some asymmetry of information. But analytical methods will not take us very far. The properties of more complex systems can only be investigated through computer simulations.

   Computable Economics will not only mean the study of more complex systems; it will also bring a different orientation towards the modelling of the elements of those systems, that is towards the representation of individual behavior.

   My friend, Daniel Heymann, once remarked that practical men of affairs, if they know anything about economics. often distrust it because it seems to describe the behavior of incredibly smart people in unbelievably simple situations. Now, non-economists often fail to understand that standard economic theory is useful in a myriad ways, despite its unrealistic assumptions about people's cognitive capabilities, because the interaction of ordinary people in markets very often does produce the incredibly smart result. When it does, it can be a convenient short-cut to model the social interaction process as if it was planned (and policed) by a representative agent or social planner possessed of rather superhuman abilities.

   The defense of our craft that I have just sketched is in the best UCLA tradition, going back to Armen Alchian's classic "Uncertainty, Evolution, and Economic Theory" [1]. But Alchian was advocating a method very much at variance with the one that dominates macrotheory today, a method

   .. which treats the decisions and criteria dictated by the economic system as more important than those made by the individuals in it. By backing away from the trees - the optimization calculus by individual units - we can better discern the forest of impersonal market forces [1, 2091.(1)

   Efficiency, in Alchian's theory, stems less from the ex ante rational planning of typical economic agents than from the ex post elimination through competition of ill-adapted modes of behavior.(2)

   We might start, then, by asking how believably simple people cope with incredibly complex situations. If we knew a bit about that, we could then go on to study the conditions under which market interaction will and will not configure the complex system into that incredibly smart allocational pattern. Because, of course, social interaction does not always produce the perfectly rational result. Sometimes, as James Tobin once said, "the invisible hand is nowhere to be seen." Ordinary people also interact to produce booms and busts in real estate, credit crunches and bank panics, great depressions and hyperinflations - and much other misery besides.

   What we should aim for is to model "systems that function pretty well most of the time but sometimes work very badly to coordinate activities" [21; 23].

   The true descendants of Alchian in more recent times are Jack Hirshleifer [16], Richard Day [8] and Nelson and Winter [30]. But standard economic theory has not taken this tack. It proceeds instead from the foundational postulate that people are rational in the sense that they will act in the manner most appropriate to the situation. Rationality is postulated here in the service of the explanatory strategy that Latsis [18] has termed "situational determinism". Situational determinism is so called because one seeks to explain or predict the behavior of an individual or organization from the external situation alone. Bounded rationality is banished in the hope that so doing will guarantee a unique prediction from the given external situation. Unbounded rationality makes "internal" questions of how decisions are reached uninteresting at best. Situational determinism, therefore, is an important fortress in the boundary defenses of economics. It has allowed the economists, most particularly, to ignore developments in the cognitive sciences. By the same token, if we give it up, the boundary defenses come down. We then have to face the frightening prospect that people in other disciplines may have something to teach us!

   Situational determinism is implemented through various optimizing techniques, where the only "internal" factor is the criterion function to be optimized. Optimizing models do not allow a natural characterization either of decision-making under incomplete knowledge (ignorance) or of adaptive dynamic processes.

   This research program has never lacked external critics. Now it has run into trouble on terms internal to itself. Problems have surfaced both on the level of individual behavior and on that of systemic behavior:

   First, certain decisions can be shown to be uncomputable (undecideable). This means, among other things, that individual behavior in such cases cannot be predicted from a description of the external situation alone. (Of course, computable economics will not solve uncomputable problems either - but it will enable us to determine where the frontier between the computable and the uncomputable runs).

   Second, dynamic general equilibrium models have been found generally to have multiple equilibria. Thus, uniqueness is not guaranteed even in large numbers (competitive) cases. Even in principle, complete information on the "fundamentals" - on tastes, technologies, and initial resource endowments - will not uniquely determine expectations and so will not suffice to determine a unique time-path for such a system. Thus, "unbounded rationality" will not buy us "situational determinism" after all. The main attraction of this unpalatable assumption is gone, therefore. To my mind, moreover, the multiple equilibria throw doubt on the behavior assumptions that produce them.

   Game theorists take to computable methods far more readily than general equilibrium and macroeconomists.(3) They are used to dealing with problems where the right equilibrium concept is not obvious and where rationality postulates will not buy you a convincing answer [5; 9]. It is also in this branch of our field that the value of studying complex dynamics on the computer is being most rapidly proven. Even so, it is notable that much of the most exciting work in computable game theory is being done by non-economists.(4) Another group that should make natural allies (for much the same reasons) are the experimentalists. Computable economics is in effect a brand of experimental economics done with artificially intelligent agents.(5)

   Bottoms Up

   Where do computers come in, you may ask? So far, I have been trying to persuade you that, in economics, we have gotten the relationship between the system and its elements - that is, between the economy and its individual agents - backwards.

   A sideways glance at what is going on in other fields can sometimes help one's perspective. The field of Artificial Intelligence is in the grips of a controversy between those who advocate a "top down" and those who favor a "bottom up" approach. The "top down" approach relies on the sheer crunching power of a centralized processor eventually to replicate whatever human intelligence can do. The "bottom up" approach, or "distributed AI", relies on interacting networks of relatively simple processors and attempts to make neural nets evolve that, by parallel processing, will handle tasks far beyond the capacities of the components.

   Neoclassical general equilibrium theory is, in these terms, quintessentially "top down". That is why, in the absence of externalities, it reduces to the optimal solution of a social planner's problem. There is little purpose to economists choosing sides and doing battle as if these two approaches were mutually exclusive across the entire discipline. But to get a handle on such ill-coordinated processes as high inflations or deep depressions, for instance, we may do better to view the system from the "bottom up." I propose the following conjecture:

   CONJECTURE. The economy is best conceived of as a network of interacting processors, each one with less capability to process information than would be required of a central processor set to solve the overall allocation problem for the entire system.

   Developments in computer science promise to be helpful in eventually implementing a research program along such lines. The first generation of parallel computers were programmed to rely on a central processor to "Gosplan" the allocation and scheduling of tasks to the subordinate processors. Recent work at Xerox laboratories [38] has been directed towards realizing truly decentralized distributed processing. Their SPAWN program operates in effect on "market" principles, making the work flow...