Optimisation of Generalised Policies via Evolutionary Computation

This paper investigates the application of Evolutionary Computation to the induction of generalised policies. A policy is here defined as a list of rules that specify which actions to be performed under which conditions. A policy is domain specific and is used in conjunction with an inference mechanism(to decide which rule to apply) to formulate plans for problems within that domain. Evolutionary Computation is concerned with the design and application of stochastic population-based iterative methods inspired by natural evolution.This work illustrates how it may be applied to the induction of policies, compares the results on one domain with those obtained by a state-of-the-art approximate policy iteration approach, and highlights both the current limitations (such as a simplistic knowledge representation) and the advantages(including optimisation of rule order within a policy) of our system.