Induction of procedures in simulated worlds

Abstract

Induction of action sequences in a simulated robot world is described. The learner passively observes examples which are procedural action sequences where properties and relations change in the environment. The complete observed procedures are then split into sub- and sub-sub- procedures with irrelevant predicates trimmed out to form the initial theory. Active experimentation occurs when the system attempts to imitate these observed examples, or plans to test more general descriptions of them. Generalisations are formed by constructive induction in two forms: absorption and intra-construction. The theory is executed and at each level a generalisation may be tested. Completed execution of the test confirms success. Inability to complete a concept in the simulation causes a regeneralisation to occur where a set of generalisations that are consistent with the failure can be found to continue execution. Background knowledge need not be present as the system invents appropriately justified concepts, as required.