Inductive Inference Research Status , Spring 1967

Previously, four theories of inductive inference were proposed, and the present work is concerned with putting these theories on a firmer theoretical footing. The four theories are now found to be effectively equivalent, at varying levels of rigor. It is shown that for sufficiently long sequences of symbols, the theories actually do give the correct probabilities for future symbols. Attempts are made to implement induction theories through a general approach to networks of tasks. Applications of induction theory to man–to–animal communication are considered, and some new approaches to interspecies communication are suggested. It is now shown that the induction theories developed for discrete symbols can be applied to prediction of continuous data. Computer implementation of induction through use of definitions of subsequences of symbols may prove to be a valuable line of effort. Supervisory problems in dealing with very intelligent machines are discussed, parallels are shown to the problem of dealing with ambitious subordinates, and a remedy for the case of the machines is proposed.