Malcolm Forster

Traditional analyses of the curve fitting problem maintain that the data do not indicate what form the fitted curve should take. Rather, this issue is said to be settled by prior probabilities, by simplicity, or by a background theory. In this paper, we describe a result due to Akaike [1973], which shows how the data can underwrite an inference concerning the curve's form based on an estimate of how predictively accurate it will be. We argue that this approach throws light on the theoretical virtues of parsimoniousness, unification, and non ad hocness, on the dispute about Bayesianism, and on empiricism and scientific realism. * Both of us gratefully acknowledge support from the Graduate School at the University of Wisconsin-Madison, and NSF grant DIR-8822278 (M.F.) and NSF grant SBE-9212294 (E.S.). Special thanks go to A. W. F. Edwards.William Harper. Martin Leckey. Brian Skyrms, and especially Peter Turney for helpful comments on an earlier draft.

Classical mechanics is empirically successful because the probabilistic mean values of quantum mechanical observables follow the classical equations of motion to a good approximation (Messiah 1970, 215). We examine this claim for the one-dimensional motion of a particle in a box, and extend the idea by deriving a special case of the ideal gas law in terms of the mean value of a generalized force used to define "pressure." The examples illustrate the importance of probabilistic averaging as a method of abstracting away from the messy details of microphenomena, not only in physics, but in other sciences as well.

The likelihood theory of evidence (LTE) says, roughly, that all the information relevant to the bearing of data on hypotheses (or models) is contained in the likelihoods. There exist counterexamples in which one can tell which of two hypotheses is true from the full data, but not from the likelihoods alone. These examples suggest that some forms of scientific reasoning, such as the consilience of inductions (Whewell, 1858. In Novum organon renovatum (Part II of the 3rd ed.). The philosophy of the inductive sciences. London: Cass, 1967), cannot be represented within Bayesian and Likelihoodist philosophies of science.

Although in every inductive inference, an act of invention is requisite, the act soon slips out of notice. Although we bind together facts by superinducing upon them a new Conception, this Conception, once introduced and applied, is looked upon as inseparably connected with the facts, and necessarily implied in them. Having once had the phenomena bound together in their minds in virtue of the Conception men can no longer easily restore them back to the detached and incoherent condition in which they were before they were thus combined. The pearls once strung, they seem to form a chain by their nature. Induction has given them unity which it is so far from costing us an effort to preserve, that it requires an effort to imagine it dissolved — William Whewell, 1858. (Quoted from Butts (ed.), 1989, p. 143).

What has science actually achieved? A theory of achievement should define what has been achieved, describe the means or methods used in science, and explain how such methods lead to such achievements. Predictive accuracy is one truth‐related achievement of science, and there is an explanation of why common scientific practices tend to increase predictive accuracy. Akaike’s explanation for the success of AIC is limited to interpolative predictive accuracy. But therein lies the strength of the general framework, for it also provides a clear formulation of many open problems of research

The simple question, what is empirical success? turns out to have a surprisingly complicated answer. We need to distinguish between meritorious fit and ‘fudged fit', which is akin to the distinction between prediction and accommodation. The final proposal is that empirical success emerges in a theory dependent way from the agreement of independent measurements of theoretically postulated quantities. Implications for realism and Bayesianism are discussed. ‡This paper was written when I was a visiting fellow at the Center for Philosophy of Science at the University of Pittsburgh; I thank everyone for their support. †To contact the author, please write to: Department of Philosophy, University of Wisconsin–Madison, 5185 Helen C. White Hall, 600 North Park Street, Madison, WI 53706; e-mail: [email protected].

Ramsey, Stick and Garon (1991) argue that if the correct theory of mind is some parallel distributed processing theory, then folk psychology must be false. Their idea is that if the nodes and connections that encode one representation are causally active then all representations encoded by the same set of nodes and connections are also causally active. We present a clear, and concrete, counterexample to RSG's argument. In conclusion, we suggest that folk psychology and connectionism are best understood as complementary theories. Each has different limitations, yet each will co-evolve with the other in an overlapping domain of 'normal' psychology.

Textbooks in quantum mechanics frequently claim that quantum mechanics explains the success of classical mechanics because “the mean values [of quantum mechanical observables] follow the classical equations of motion to a good approximation,” while “the dimensions of the wave packet be small with respect to the characteristic dimensions of the problem.” The equations in question are Ehrenfest’s famous equations. We examine this case for the one-dimensional motion of a particle in a box, and extend the idea deriving a special case of the ideal gas law in terms of the mean value of a generalized force, which has been used in statistical mechanics to define ‘pressure’. The example may be an important test case for recent philosophical theories about the relationship between micro-theories and macro-theories in science.

The theory of fast and frugal heuristics, developed in a new book called Simple Heuristics that make Us Smart (Gigerenzer, Todd, and the ABC Research Group, in press), includes two requirements for rational decision making. One is that decision rules are bounded in their rationality –- that rules are frugal in what they take into account, and therefore fast in their operation. The second is that the rules are ecologically adapted to the environment, which means that they `fit to reality.' The main purpose of this article is to apply these ideas to learning rules–-methods for constructing, selecting, or evaluating competing hypotheses in science, and to the methodology of machine learning, of which connectionist learning is a special case. The bad news is that ecological validity is particularly difficult to implement and difficult to understand. The good news is that it builds an important bridge from normative psychology and machine learning to recent work in the philosophy of science, which considers predictive accuracy to be a primary goal of science.

Skyrms's formulation of the argument against stochastic hidden variables in quantum mechanics using conditionals with chance consequences suffers from an ambiguity in its "conservation" assumption. The strong version, which Skyrms needs, packs in a "no-rapport" assumption in addition to the weaker statement of the "experimental facts." On the positive side, I argue that Skyrms's proof has two unnoted virtues (not shared by previous proofs): (1) it shows that certain difficulties that arise for deterministic hidden variable theories that exploit a nonclassical probability theory extend to the stochastic case; (2) the use of counterfactual conditionals relates the Bell puzzle to Dummett's (1976) discussion of realism in quantum mechanics

Van Fraassen has argued that quantum mechanics does not conform to the pattern of common cause explanation used by Salmon as a precise formulation of Smart's 'cosmic coincidence' argument for scientific realism. This paper adds to this list some common examples from classical physics that also do not conform to Salmon's explanatory schema. This is bad news and good news for the realist. The bad news is that Salmon's argument for realism does not work; the good news is that realism need not demand hidden variables in quantum mechanics if they are not used in classical mechanics. Many correlations in physics are explained in terms of property identity (contra Salmon). This leads to a new argument against van Fraassen because the unified version of the theory obtained by identifying theoretical properties is always less empirically adequate

Recent approaches to causal modelling rely upon the causal Markov condition, which specifies which probability distributions are compatible with a directed acyclic graph. Further principles are required in order to choose among the large number of DAGs compatible with a given probability distribution. Here we present a principle that we call frugality. This principle tells one to choose the DAG with the fewest causal arrows. We argue that frugality has several desirable properties compared to the other principles that have been suggested, including the well-known causal faithfulness condition. _1_ Introduction _2_ The Causal Markov Condition _3_ Faithfulness _4_ Frugality _4.1_ Basic independences and frugality _4.2_ General properties of directed acyclic graphs satisfying frugality _4.3_ Connection to minimality assumptions _5_ Frugality as a Parsimony Principle _6_ Conclusion Appendix

What is induction? John Stuart Mill (1874, p. 208) defined induction as the operation of discovering and proving general propositions. William Whewell (in Butts, 1989, p. 266) agrees with Mill’s definition as far as it goes. Is Whewell therefore assenting to the standard concept of induction, which talks of inferring a generalization of the form “All As are Bs” from the premise that “All observed As are Bs”? Does Whewell agree, to use Mill’s example, that inferring “All humans are mortal” from the premise that “John, Peter and Paul, etc., are mortal” is an example of induction? The surprising answer is “no”. How can this be?

Recent solutions to the curve-fitting problem, described in Forster and Sober ([1995]), trade off the simplicity and fit of hypotheses by defining simplicity as the paucity of adjustable parameters. Scott De Vito ([1997]) charges that these solutions are 'conventional' because he thinks that the number of adjustable parameters may change when the hypotheses are described differently. This he believes is exactly what is illustrated in Goodman's new riddle of induction, otherwise known as the grue problem. However, the 'number of adjustable parameters' is actually a loose way of referring to a quantity that is not language dependent. The quantity arises out of Akaike's theorem in a way that ensures its language invariance.

William Whewell’s philosophy of scientific discovery is applied to the problem of understanding the nature of unification and explanation by the composition of causes in Newtonian mechanics. The essay attempts to demonstrate: the sense in which ”approximate’ laws successfully refer to real physical systems rather than to idealizations of them; why good theoretical constructs are not badly underdetermined by observation; and why, in particular, Newtonian forces are not conventional and how empiricist arguments against the existence of component causes, and against the veracity of the fundamental laws, are flawed.

The paper provides a formal proof that efficient estimates of parameters, which vary as as little as possible when measurements are repeated, may be expected to provide more accurate predictions. The definition of predictive accuracy is motivated by the work of Akaike (1973). Surprisingly, the same explanation provides a novel solution for a well known problem for standard theories of scientific confirmation — the Ravens Paradox. This is significant in light of the fact that standard Bayesian analyses of the paradox fail to account for the predictive utility of universal laws like All ravens are black.

Deductive logic is about the validity of arguments. An argument is valid when its conclusion follows deductively from its premises. Here’s an example: If Alice is guilty then Bob is guilty, and Alice is guilty. Therefore, Bob is guilty. The validity of the argument has nothing to do with what the argument is about. It has nothing to do with the meaning, or content, of the argument beyond the meaning of logical phrases such as if…then. Thus, any argument of the following form (called modus ponens) is valid: If P then Q, and P, therefore Q. Any claims substituted for P and Q lead to an argument that is valid. Probability theory is also content-free in the same sense. This is why deductive logic and probability theory have traditionally been the main technical tools in philosophy of science.

Scientific discovery is often regarded as romantic and creative - and hence unanalyzable - whereas the everyday process of verifying discoveries is sober and more suited to analysis. Yet this fascinating exploration of how scientific work proceeds argues that however sudden the moment of discovery may seem, the discovery process can be described and modeled. Using the methods and concepts of contemporary information-processing psychology (or cognitive science) the authors develop a series of artificial-intelligence programs that can simulate the human thought processes used to discover scientific laws. The programs - BACON, DALTON, GLAUBER, and STAHL - are all largely data-driven, that is, when presented with series of chemical or physical measurements they search for uniformities and linking elements, generating and checking hypotheses and creating new concepts as they go along. "Scientific Discovery examines the nature of scientific research and reviews the arguments for and against a normative theory of discovery; describes the evolution of the BACON programs, which discover quantitative empirical laws and invent new concepts; presents programs that discover laws in qualitative and quantitative data; and ties the results together, suggesting how a combined and extended program might find research problems, invent new instruments, and invent appropriate problem representations. Numerous prominent historical examples of discoveries from physics and chemistry are used as tests for the programs and anchor the discussion concretely in the history of science. Pat Langley is an Associate Professor in the Department of Information and Computer Science at the University of California, Irvine.Herbert Simon is a Professor in the Departments of Psychology, Computer Science, and Philosophy at Carnegie-Mellon University. Gary L. Bradshaw is an Assistant Professor in the Department of Psychology and Institute of Cognitive Science at the University of Colorado, Boulder. Jan M. Zytkow is an Associate Professor in the Computer Science Department at Wichita State University.

Sober (1984) has considered the problem of determining the evidential support, in terms of likelihood, for a hypothesis that is incomplete in the sense of not providing a unique probability function over the event space in its domain. Causal hypotheses are typically like this because they do not specify the probability of their initial conditions. Sober's (1984) solution to this problem does not work, as will be shown by examining his own biological examples of common cause explanation. The proposed solution will lead to the conclusion, contra Sober, that common cause hypotheses explain statistical correlations and not matchings between event tokens

Kenneth Wilson won the Nobel Prize in Physics in 1982 for applying renormalization group, which he learnt from quantum field theory (QFT), to problems in statistical physics—the induced magnetization of materials (ferromagnetism) and the evaporation and condensation of fluids (phase transitions). See Wilson (1983). The renormalization group got its name from its early applications in QFT. There, it appeared to be a rather ad hoc method of subtracting away unwanted infinities. The further allegation was that the procedure is so horrendously complicated that one cannot see the forest for the trees. The second allegation is justified in the applications that made it famous. But it is not true of the following example, which appears in Chowdhury and Stauffer (2000, 486-488).

The Value of Good Illustrative Examples: In order to speak as generally as possible about science, philosophers of science have traditionally formulated their theses in terms of elementary logic and elementary probability theory. They often point to real scientific examples without explaining them in detail and/or use artificial examples that fail to fit with intricacies of real examples. Sometimes their illustrative examples are chosen to fit their framework, rather than the science. Frequently these are non-scientific examples, which distances the discussion from its intended target. In the final analysis, philosophical discussions of explanation, confirmation, scientific realism, and the nature of theories are often too abstract, or too imprecise, or too disconnected with real science, to allow scientists to benefit from the discussion. This is a great loss for both parties. In my experience, working scientists are confronted with philosophical issues not only in their role as researchers, but also in their role as tertiary teachers of science.

What has science actually achieved? A theory of achievement should define what has been achieved, describe the means or methods used in science, and explain how such methods lead to such achievements. Predictive accuracy is one truth-related achievement of science, and there is an explanation of why common scientific practices tend to increase predictive accuracy. Akaike's explanation for the success of AIC is limited to interpolative predictive accuracy. But therein lies the strength of the general framework, for it also provides a clear formulation of many open problems of research.