In my discussion of Sellars I have said as little as possible about my own position. But I think I should just mention that the descriptions of instrumentalists by Professors Sellars and Cornman do not fit me. That is why I would prefer the term “anti-realist”.

To my mind, theoretical entities are fictions. To explain this, let me draw an analogy. Suppose someone writes a short story about a quarrel between a man and a cat. There may be or have been in the world, somewhere, a man and a cat who quarrelled, and who are by and large just like the characters in the story. This possibility is quite irrelevant to what the author is doing, and to our evaluation of the story. Of course, it keeps us from saying categorically that all short stories are false; but this too is quite irrelevant.

In his paper prepared for the 1976 meeting of the Philosophy of Science Association, Professor Sellars has raised two quite different objections, one against Professor van Fraassen and one against me. Although it may be that I should limit my reply to the issue concerning my theory of perceiving that Sellars raised, I cannot resist injecting myself into the debate between Sellars and van Fraassen. One reason I want to engage in their battle is that I find reasons for disagreeing with both of them on certain important issues which they have debated. My hope is that by joining their discussion I will help clarify the issues involved. After that, however, I will reply to Sellars’ objection to my theory of perceptual experience and perceiving.

Sellars and van Fraassen are diametrically opposed on the issue of scientific realism versus scientific instrumentalism.

The foregoing is the only example that L. J. Savage uses in the book that made Bayesian decision theory respectable. In discussing this example, Savage made a very interesting claim, that I want to explore here. The exploration will be far from complete.

Even a distant glance reveals that in philosophy (as in my own field of economics) it is sometimes best to develop arguments at a very general and abstract level. There are no doubt a great many propositions in logic, epistemology, and the theory of probability, for example, which are relevant to many diverse activities and aspects of reality. If we did not abstract from the peculiarities of particular cases we presumably could not so readily illuminate certain logical or conceptual problems nor so tersely express our findings. This is unquestionably true in economics: there are many important propositions in microeconomic theory, for example, that apply about as well to one industry or society as to another, and both the development and teaching of the theory would be slowed down if highly abstract arguments were to be ruled out.

What this paper has to do with the theory of rational choice is not clear. Still, I won't apologize. The ideas here grow out of a study of some applications of rationalism. It seems to me that the rationalist model neglects a great deal. This is not a criticism. Every model abstracts from nature. But the aspects of choosing what this model ignores are often those we most need to note in order to explain behavior. And we need to bring them out for other purposes too. Hence this essay toward a fuller analysis.

First about rational choice theory itself. On the rationalist theory, every person pursues his interests, and, in particular, pursues them in a certain way. He considers the consequences of his options, and chooses that option whose consequences he prefers to the consequences of each of the others.

One of the more important directions in which work in decision theory is moving is towards the consideration of possible norms not only upon the choices of actions but also upon preferences among the outcomes. There is beginning to be felt a need to explore rationality conditions upon desires, or goals, or preferences, in the same way that rationality conditions upon beliefs and upon choices of actions have been explored. I have suggested on another occasion that there are also reasons to explore the possibility of norms on the processes of planning, design of alternatives, and adaptation. I will not be dealing with these latter issues in this paper, however. Neither do I want now to discuss at length the reasons for incorporating norms on goals, etc., into a decision theory. Yet perhaps I need to motivate the project somewhat at the outset.

Inductive logic studies the structural properties of epistemic probabilities. Its basic task is to determine how various factors may influence rational degrees of belief in the truth of hypotheses on the basis of some evidence. Its distinctive feature in comparison with other approaches to epistemic probabilities is the attention paid to the idea - not in vogue during the 1940's when Carnap started his program, but very influential in much of the contemporary work within the history and the philosophy of science - that our knowledge claims, or our rational beliefs about the reality, are significantly influenced or preconditioned. by language. While dependence on (formal) language is one of the “logical” aspects of inductive probabilities, the role of other relevant, possibly extra-logical, determinants of rational degrees of belief is expressed, within the systems of inductive logic, by means of at most a finite number of parameters.

I have recently developed two kinds of model for inductive logic with causal modalities, the deterministic and the probabilistic. Roughly speaking, the first provides certain syntactic and semantic criteria for causal necessity, so that a causal or nomic universal can be deductively derived from a certain actual universal; whereas the second model allows only a probabilistic transition from actual universality to causal universality. These models are meant to provide, from a modal point of view, a formal explication of nomic universality both in its semantic and inductive aspects.

However, these models are too abstract if we want to treat adequately such problems as causality, the confirmability of causal laws, etc., in that they do not have a framework of space and time, which seems essential to our conception of causality.

One of the salient features of Carnap's systems of inductive logic is a conditionalization learning model, which also plays a fundamental role in the orthodox Bayesian account of rationality. This learning model does not allow for revision of previously accepted evidence. It is, therefore, not adequate to represent all rational learning by an agent who accepts corrigible propositions as evidence. Recently I used a generalization of the concept of conditional belief to extend the model so that rational revision of previously accepted evidence can be accommodated. My generalization of conditional belief carries with it a natural way of representing an agent's conceptual framework. In this paper I exploit this representation to produce learning models that can accommodate rational conceptual change.

The idea of conceptual change has received considerable attention in recent work of philosophers of science. The following quotation from Hilary Putnam is instructive.

As it has turned out—and not, I think, by conscious design—this symposium has a truly Hegelian structure. The thesis and antithesis are provided by the themes of complexity and simplicity, respectively. Professor Nelson pleads the case for complexity, and I for simplicity; and we both count on Professor Suppes to propose the synthesis.

Methods of characterizing the simplicity or complexity of systems are many. Let me mention some of them:

• (1) Systems that have many components may be considered complex relative to systems that have few. Hence the cardinality of a set may be taken as one measure of its complexity.

• (2) Systems in which there is much interdependence among the components are generally regarded as more complex than systems with less interdependence among components.

• (3) Systems that are undecidable may be regarded as complex in comparison with those that are decidable.

The area of fuzzy complex systems is a good one for analytic philosophers to keep out of. If analysis is a way of getting straight on ideas or terms which enjoy wide philosophic and scientific currency and which have fairly well-fixed meanings awaiting illumination, then complex systems is no field for analysis. I have tried it; others have tried it; and it seems to me that the attempts amount to very little. Perhaps Professor Forrester and his MIT colleagues would say that positive feedback is at work. Just as expanded, low cost housing aggravates the housing situation in large cities, so attempts to shed philosophy's light on very large scale systems, fuzzy or not, brings darkness ([3], pp. 65ff.).

There appear to be two types of systems concept that are bothersome, one benign and the other not.

My remarks on Nelson and Simon are directed not so much at criticisms of what they have had to say as at questions generated by reading and studying what they have written. I also emphasize that my questions are genuine questions. They are not a rhetorical device for stating my own clear and definite views on the nature of complex systems. I have organized these remarks under seven headings.

Although the official title of our symposium is “Complex Fuzzy Systems,” neither Nelson nor Simon spends any time on what it means to say that a system is not merely a complex system but a complex fuzzy system. Presumably, the organizers of the symposium had in mind the concept of fuzzy introduced by Zadeh.

Statistical mechanics, in its broad sense, means the study of properties of bulk matter, assuming it to be composed of large numbers of particles. While each particle obeys the laws of mechanics — classical, quantum, or relativistic, whichever may be appropriate — the application of statistical methods avoids the need for considering the motion of each particle in detail, and suggests that the system as a whole has certain simple regularities. Many of these regularities were first explained (or even predicted) with the help of the kinetic theory of gases, an early version of statistical mechanics in which intuitive physical ideas based on special molecular models were applied to systems of particles whose encounters are relatively infrequent. The more general theory, developed by Maxwell, Boltzmann, and Gibbs, attempts to deal with a broad class of mechanical systems, using consistent statistical postulates.

The idea of combining molecular mechanics with statistical principles, in order to provide a “reduction” of thermodynamics to mechanics, took shape about a hundred years ago. Since that time, the technique of explaining macroscopic properties of matter in microscopic terms has advanced by leaps and bounds; but this practical success is not matched by a comparable progress in our understanding of the “foundations”. Recent mathematical advances in the ergodic theory ([18] and the various papers published in [6]) make it only more apparent that such discussions are still pursued in the conceptual framework of Boltzmann. My purpose is to advance another approach that is in some sense complementary to the traditional one. I suggest first a heuristic guideline that is broad enough to subsume both programs.

Most of what I have to say concerns Tisza's conceptual analysis [12], but I should like to begin with an appreciation of Brush's historical essay [2], Accurate historical information about a science is usually illuminating not only about the process of discovery but also about the subject matter itself. I think that this statement is especially true for statistical mechanics, in which the problems that the founding fathers were concerned with are still not completely solved, and I certainly feel that this historical essay has helped me to understand the subject matter of statistical mechanics better than before.

There are several remarks in Brush's essay concerning epistemology and methodology, some of which seem to me correct but others not.