Multi-Task Multi-Agent Reinforcement Learning With Task-Entity Transformers and Value Decomposition Training