Computational Measures of Information Gain and Reinforcement in Inference Processes


Author:
Title:
Language of presentation:
Supervisor:
Date of defence:
Institution granting degree:
José Hernández-Orallo
Computational Measures of Information Gain and Reinforcement in Inference Processes
English
Prof. Rafael Beneyto
September 1996
Universitat de València, Spain

Abstract

This work is devoted to the formal study of inductive and deductive concept synthesis usefulness and aftermath in terms of information gain and reinforcement inside inference systems. The set of measures which are introduced allow a detailed and unified analysis of the value of the output of any inference process with respect to the input and the context (background knowledge or axiomatic system).

Although the main measures, computational information gain, reinforcement and intensionality, are defined independently, they (alone or combined) make it possible to formalise or better comprehend several notions which have been traditionally treated in a rather ambiguous way: novelty, explicitness/implicitness, informativeness, surprise, interestingness, plausibility, confirmation, comprehensibility, ‘consilience’, utility and unquestionability.

Most of the measures are applied to different kinds of theories and systems, from the appraisal of predictiveness, the representational optimality and the axiomatic power of logical theories, software systems and databases, to the justified evaluation of the intellectual abilities of cognitive agents and human beings.