Sample efficiency analysis of Neuroevolution algorithms on a quadruped robot

In reinforcement learning tasks with continuous state-action, parameterized policy search has been known to be a powerful method. Applying NeuroEvolution (NE) to optimizing the policy represented by artificial neural network (ANN) is a particularly active research field. In most cases, NE algorithms cost a large amount of trial-and-error (episode) to optimize policies. However, due to time and cost constraints, researchers and practitioners cannot repeat a number of episodes on physical robots. Thus, choosing an efficient NE algorithm is a key to optimize policies with limited time and cost. In this work, our goal is to help users to choose an efficient NE algorithm. We compare and analyze sample efficiency of two successful state-of-the-art NE algorithms: CMA-NeuroES and NEAT in a gait generation task of a quadruped robot. Moreover, we run both algorithms with various initial topologies in order to analyze the performance difference between each topology. From experimental results, we show CMA-NeuroES outperforms NEAT regardless of initial topologies when the limited number of episodes can be executed. Additional experiments conclude that the optimization method for connection weights in NEAT results in its inferior performance to CMA-NeuroES, while a probability-weighted averaging characteristic and self-adaptive factors make CMA-NeuroES to be advantageous.