Problem decomposition and the learning of skills

Abstract

One dimension of “divide and conquer” in problem solving concerns the domain and its subdomains. Humans learn the general structure of a domain while solving particular learning problems in it. Another dimension concerns the solver's goals and subgoals. Finding good decompositions is a major AI tactic both for defusing the combinatorial explosion and for ensuring a transparent end-product. In machine learning, pre-occupation with free-standing performance has led to comparative neglect of this resource, illustrated under the following headings. 1. Automatic manufacture of new attributes from primitives (“constructive induction”). 2. Machine learning within goal-subgoal hierarchies (“structured induction”). 3. Reconstruction of skills from human performance data (“behavioural cloning”).