Prof. Andrew Lim
National University of Singapore & University of California Berkeley
Talk:
Decentralized control of a stochastic multi-agent service system
Abstract:
Consider a stochastic decentralized multi-agent service system consisting of a pricing and a service agent, where the pricing agent dynamically controls the customer arrival rate and the service agent controls the rate at which customers are processed. With the aid of transfer/revenue sharing contracts, we show that this integrated problem can be decoupled into a dynamic pricing problem for the pricing agents and a service rate control problem for the service agent, and that contracts can be specified such that decentralized agent level optimization delivers centrally optimal policies. This property is shown to be robust to mis-specification by each of the agents of the impact of other agent on the system and does not require any of the agents to reveal private information (demand models, etc). An iterative algorithm for computing centralized contracts will also be presented and shown to converge to the optimal contract. A key feature of this algorithm is that its implementation does not require agents to reveal private information. Time permitting, extensions to other stochastic dynamic resource allocation problems, such as dynamic optimal portfolio selection with heterogeneous decentralized agents and coordination of airline alliances, will also be discussed.
Biography:
Andrew Lim joined the IEOR department at UC Berkeley in 2002. His research and teaching interests include stochastic control and optimization, financial engineering and risk management. Much of his recent work has focused on the problem of decision making in stochastic settings when there is model uncertainty.
