Action Semantics Network: Considering the Effects of Actions in Multiagent Systems

Published in Proceedings of International Conference on Learning Representation (ICLR), 2020


This paper proposes a novel network architecture, named Action Semantics Network (ASN), that explicitly characterizes different actions’ influence on other agents using neural networks based on the action semantics between them. Experimental results on StarCraft II micromanagement and Neural MMO show ASN significantly improves the performance of state-of-the-art DRL approaches compared with several network architectures.