Attention-Based Curiosity in Multi-Agent Reinforcement Learning Environments

Abstract

Several paradigms exist in Reinforcement Learning to improve the exploration capabilities of agents, among which the curiosity-driven approach is followed in this work. Extending previous work that utilizes attention to make curiosity state-and action-selective, we expand the range of experiments by introducing two multi-agent environments. The first one is for robot soccer, the second one features a driving scenario in urban settings. Moreover, as during training the different number of observations must be matched between multiple time-steps, we propose an attention-based approach, called Recurrent Temporal Attention (RTA) to do this. The corresponding implementation can be found at https://github.com/szemenyeim/DynEnv.