My research objective is to develop scalable solvers for planning and scheduling problems in multi-agent systems. These include but are not limited to
Distributed computation In many real applications, centralized control is not available but it requires autonomous agents to coordinate in a decentralized system, e.g. rescue robots in disaster, ad-hoc sensors with limited communication. In such decentralized setting, each agent needs to find policy for its local viewpoint to maximize not only its observed reward but also the global welfare. I study distributed algorithm to specify how each agent should exchange the information when computing its local action.
Mechanism design for noncooperative agents Individual benefit could conflict with social welfare in many cases. Greediness causes rational agents to choose the best action to maximize individual rewards, disregarding to others'. This can lead to non-agreement between centralized solution maximizing the global utility of the system and individual solution maximizing local reward. To resolve the potential conflict, I study incentive compatible solution which is "acceptable" in each of individual agent perspective when each agent could not improve its local reward further. My motivating domain is taxi-sharing systems in which passengers expect to share taxis with different willing payment. Regard to the heterogeneity of passenger willing payments, we study incentive compatible mechanisms to provide compact solution determining both vehicle routing and cost allocating.
Collective inference and planning In contrast to simplified multi-agent computational model with a tractable number of dozens or hundreds of agents, in many real-life domains we have to deal with a large population of thousands or even million of agents. For example, coordinating vehicles in a traffic network to reduce the congestion or dispatching a fleet of taxis to match passengers better. It is impossible to keep track of every single agent. In such domain, the identity of an agent can be marginalized out into aggregated statistics, e.g. counting the number of agents in each type or at each state, which reduces the complexity of the problem. Our goal is to establish scalable collective solvers for inference and planning tasks based on the aggregate states.