Efficient Neuroevolution for a Quadruped Robot

摘要

In this research, we investigate whether CoSyNE and CMA-NeuroES algorithms can efficiently optimize neural policy of a quadruped robot. Both of these algorithms are proven to optimize connection weights efficiently on Pole Balancing benchmark. Due to their good results on that benchmark, they are expected to be efficient on other control problems like gait generation. In this research we experimentally show that CMA-NeuroES have higher scalability to optimize Artificial Neural Networks for generating gaits of quadruped robots in comparison with CoSyNE. The results can be helpful for researchers and practitioners to choose the optimal Neuroevolution algorithm for generating gaits.