ROBOT LEARNING WITH PARALLEL GENETIC ALGORITHMS ON NETWORKED COMPUTERS

This work explores the use of machine learning methods for extracting knowledge from simulations of complex systems. In particular, we use genetic algorithms to learn rule-based strategies used by autonomous robots. The evaluation of a given strategy may require several executions of a simulation to produce a meaningful estimate of the quality of the strategy. As a consequence, the evaluation of a single individual in the genetic algorithm requires a fairly substantial amount of computation. Such a system suggests the sort of large-grained parallelism that is available on a network of workstations. We describe an implementation of a parallel genetic algorithm, and present case studies of the resulting speedup on two robot learning tasks.