Sample Efficiency Improvement on Neuroevolution v ia Estimation-Based Elimination Strategy (Extended Abstract)

In this paper, we propose estimation-based elimination strat- egy, which improves sample efficiency of NeuroEvolution (NE) algorithms. The fitness of new individuals was estimated us- ing fitness of individuals evaluated in the past generations. The estimation was achieved by taking average fitness of individuals with high correlation with the new individual. Estimation-based elimination strategy avoids evaluating in- dividuals with low estimated fitness. We adapt estimation- based elimination strategy for state-of-the-art NE algorithms: CMA-NeuroES and CMA-TWEANN. From the experimen- tal results of pole-balancing benchmark tasks, we show that the proposed strategy improves sample efficiency of the NE algorithms.