UMass Machine Learning and Friends Lunch | Main / Towards Conquering Uncertainty In Agent Systems

In most domains involving either individual agents or teams, ranging from autonomous space exploration to automated personal assistants to sensor networks, intelligent agents must fundamentally reason with uncertainty. In providing techniques for reasoning with such uncertainty, we must address at least three major challenges: (i) these techniques handle both discrete and continuous features of the environment (ii) they are applicable to partially observable environments and (ii) they scale up in the number of agents, yet provide quality guarantees, as ad-hoc approaches without guarantees may cause significant failures in serious applications. In this talk I will present two key ideas that help to address these challenges: (i) An analytic algorithm for solving Markov Decision Processes (MDPs) with continuous resources that is orders of magnitude faster than its competitors and (ii) An algorithm for solving Distributed, Partially Observable MDPs that exploits agent interaction structure and varies agent policy expressivity to scale up the number of agents to double digits. Furthermore, I will demonstrate the practical benefit of my algorithms on the deployed systems and outline my vision of the Agent Systems of tomorrow.