Hybrid Policy Learning for Multi-Agent Pathfinding

In this work we study the behavior of groups autonomous vehicles, which are part Internet Vehicles systems. One challenging modes operation such systems is case when observability each vehicle limited and global/local communication unstable, e.g. in crowded parking lots. scenarios vehicles have to rely on local observations exhibit cooperative ensure safe efficient trips. This type problems can be abstracted so-called multi-agent pathfinding a group agents, confined graph, find collision-free paths their goals (ideally, minimizing an objective function travel time). Widely used algorithms for solving problem assumption that central controller exists full state environment (i.e. agents current positions, targets, configuration static obstacles etc.) known they not straightforwardly adapted partially-observable setups. To end, suggest novel approach based decomposition into two sub-tasks: reaching goal avoiding collisions. accomplish task utilize reinforcement learning methods as Deep Monte Carlo Tree Search, Q-mixing networks, policy gradients design policies map agents’ actions. Next, introduce policy-mixing mechanism end up with single hybrid allows agent both types – individual one (reaching goal) (avoiding collisions other agents). We conduct extensive empirical evaluation shows suggested hybrid-policy outperforms standalone stat-of-the-art kind by notable margin.