Transformer-based Working Memory for Multiagent Reinforcement Learning with Action Parsing Y aodong Y ang

Our second contribution is to introduce an action semantic inductive bias and implement it with a technique named Entity-Bound Action Layer.