Optimum inductive methods (Q1892403)

This book attempts to do two things. First to present an analytical review of two important schools of thought concerning induction: one stems from Carnap's work on the theory of inductive probability (``TIP'' in the heavily acronymic style of the author), and the other stems from the Popperian claim that science seeks ``verisimilitude'' rather than truth. Second, the author offers his own solution to ``EPO'' -- the epistemic problem of optimality. The author is extremely successful at the first task; he is less successful at the second. The basic features of inductive probability, as envisaged by Carnap, are set forth in the second chapter of the book. The focus of the discussion is the case of multinomial inference as captured by multivariate Bernoullian processes. This is general enough to serve to illustrate many distinctions and applications; but also simple enough to be readily related to its Carnapian roots. Bayesian statistics and verisimilitude theory (``VT'') are discussed in the third and fourth chapters. The second part of the book, comprising chapters five and six, is a very useful survey of results relating exchangeability, convergence, generalized Carnapian systems, and Dirichlet distributions. It is shown that a common Bayesian prior for multinomial inference and Carnap's continuum of inductive methods are both special cases of the application of Dirichlet distributions of the form: \[ \text{Dir} (q_1, \dots, q_{k-1}; \alpha_1, \dots, \alpha_k)= {{\Gamma (\alpha)} \over {\Gamma (\alpha_1) \dots \Gamma(\alpha_{k-1})}} q_1^{\alpha_1 -1} \dots q_k^{\alpha_k -1} \] where \(\alpha= \sum (\alpha_i)\) and \(q_k= 1-q_1- \dots- q_{k-1}\). With \(\alpha_1= \alpha_j\) we have the Carnapian continuum; allowing for inequalities, we have Generalized Carnapian (GC) systems. The third and final part of the book concerns the choice of prior probabilities. The author deals carefully with the distinctions among the a priori approaches (as represented by `objective Bayesians') and one strand of Carnap, the subjective approach, represented by Savage, de Finetti, and another strand of Carnap, and what he calls the `hyperempirical' approach, represented by David Lewis and Kuipers. To this he adds his own `contextual' view of prior probabilities. The contextual view makes use of the assumption that `the researcher' has available as part of his background knowledge either some experimental information about processes similar to the one under investigation, or some theoretical information about processes similar to the one under investigation. According to the author this allows the analysis in terms of verisimilitude (Popper's approach) to coincide with the analysis based on Carnap's continuum of inductive methods. The difficulty with this rosy view is that it shares with verisimilitude in general a crucially debilitating vagueness. We may grant that verisimilitude can provide insights into the general practice of science without taking it to yield a formal method of the kind that Carnap sought. But to ask for precision in the form of agreed-upon similarities is more than verisimilitude can deliver. Furthermore, if we have that agreed upon background knowledge, it is hard to see why it cannot be employed directly to support the acceptance of prior distributions, without the need of the rules of thumb (use an average, use a weighted average) suggested. Finally, the `optimality' of the Dirichlet distribution arrived at is entirely dependent on the values of the parameters assigned to the `similar' processes. A further factor to whose value we are guided by our knowledge of similar processes is the diversity of the domain we are investigating. From the beginning of the book this is measured by Gini diversity \((1-\sum q_i^2)\); one of the unexpected useful features of the book is the general treatment in chapter 10 of various measures of homogeneity and diversity both in theoretical contexts and, even more interesting, in concrete empirical applications. To say that the final result is less than persuasive is not at all to say that the course of the analysis has been uninformative. Not only is the argument clear, even when it is not persuasive, but it is based on a broad and careful reading of the major contributors to inductive logic in its broadest form. It is a shining example of careful and thorough scholarshop. Carnap would have been pleased to see his \(\lambda\)- continuum made such good use of. It also reveals a collection of solid relations among formal concepts that are seldom brought together. It will be a great value to those concerned with inductive and uncertain inference, whether in artificial intelligence, in statistical modelling, or in any of several empirical domains.