Abmarl: Connecting Agent-Based Simulations with Multi-Agent Reinforcement Learning
Edward Rusu, Ruben Glatt · The Journal of Open Source Software · 2021
Abmarl is a package for developing Agent-Based Simulations and training them with Multi-Agent Reinforcement Learning (MARL).We provide an intuitive command line interface for engaging with the full workflow of MARL experimentation: training, visualizing, and analyzing agent behavior.We define an Agent-Based Simulation Interface and Simulation Manager, which control which agents interact with the simulation at each step.We support integration with popular reinforcement learning simulation interfaces, including gym.Env (Brockman et al., 2016) and MultiAgentEnv (Liang et al., 2018).We leverage RLlib's framework for reinforcement learning and extend it to more easily support custom simulations, algorithms, and policies.We enable researchers to rapidly prototype MARL experiments and simulation design and lower the barrier for pre-existing projects to prototype Reinforcement Learning (RL) as a potential solution.