Multitask Neuroevolution for Reinforcement Learning with Long and Short Episodes

Studies have shown evolution strategies (ES) to be a promising approach for reinforcement learning (RL) with deep neural networks. However, the issue of high sample complexity persists in applications ES RL over long horizons. This paper is first address shortcoming todays methods via novel neuroevolutionary multitasking (NuEMT) algorithm, designed transfer information from set auxiliary tasks (of short episode length) target (full task at hand. The tasks, extracted target, allow an agent update and quickly evaluate policies on shorter time evolved skills are then transferred guide longer harder towards optimal policy. We demonstrate that NuEMT algorithm achieves data-efficient evolutionary RL, reducing expensive agent-environment interaction data requirements. Our key algorithmic contribution this setting introduce, time, multitask mechanism based statistical importance sampling technique. In addition, adaptive resource allocation strategy utilized assign computational resources their gleaned usefulness. Experiments range continuous control OpenAI Gym confirm our proposed efficient compared recent baselines.