From Knoesis wiki
Jump to: navigation, search

We live in a world with multiple agents (e.g., robots, web applications, people), and multi-agent reinforcement learning (MARL) is a promising framework for building intelligent agents that are successful in that world. Despite recent notable success, current works on MARL still fall short because of the following two challenges: (i) existing algorithms for learning communication heavily depend on handcrafted inductive biases, which often induces excessive communication and heavily relies on hyperparameter tuning and/or hand designed heuristics; (ii) current methods for MARL largely do not support generalization across environments, which forces us to train every MARL task from scratch. We expect our proposed framework to discover state-of-the-art MARL algorithms by which agents can be trained to effectively perform tasks using selective agent-to-agent communication that induces low overhead. The discovered MARL algorithms can be readily applied to real-world cooperative multi-agent systems, such as multi-robot systems, smart grid control, traffic control, etc.