Breeding Schedules Improve Grid Robot Performance

Abstract

As new species arise from biological evolution, new techniques arise when an evolutionary algorithm is used to train virtual robots. This study generalizes the ISAc list encoding for virtual robots working on the Tartarus task and tests a biologically inspired algorithmic technique called hybridization. A collection of 100 populations of virtual robots is trained in three distinct ways. A baseline experiment performs 1000 generations of evolution on each population. Two other collections of runs stop at intermediate points, harvest the currently best controllers from each population, and initialize new runs with the harvested controllers from all of the populations. This harvesting and mixing of evolutionary lines is called hybridization. In one experiment hybridization is performed at generation 500; the other performs it at generations 250, 500, and 750. The goal is to assess the utility of hybridization for the Tartarus task and to compare more and less frequent use of hybridization. Hybridization is found to substantially improve the training of the Tartarus agents, and more frequent hybridization yields greater benefits. The study reports a new maximum cross-validated fitness for Tartarus of 8.49.