PALO bounds for reinforcement learning in partially observable stochastic games

We introduce reinforcement learning for heterogeneous teams in which rewards an agent are additively factored into local costs, stimuli unique to each agent, and global rewards, those shared by all agents the domain. Motivating domains include coordination of varied robotic platforms, incur different costs same action, but share overall goal. present two templates this setting with rewards: a generalization Perkins' Monte Carlo exploring starts POMDPs canonical MPOMDPs, single policy mapping joint observations actions (MCES-MP); another individually their own action (MCES-FMP). use probably approximately optimal (PALO) bounds analyze sample complexity, instantiating these PALO learning. promote efficiency including space pruning technique, evaluate approaches on three demonstrating that MCES-FMP yields improved policies less samples compared MCES-MP previous benchmark.