 |
Center for Nonlinear Studies |
Compared to typical single-agent decision problems, general sum games offer a panoply of strategies for maximizing utility. In many games, such as the well-known Prisoner's dilemma, agents must work together, bearing some individual risk, to arrive at mutually beneficial outcomes. In this talk, I will discuss three algorithmic approaches that we have developed to identify cooperative strategies in non-cooperative games. I will describe a computational folk theorem, an analysis of value-function-based reinforcement learning, and a cognitive hierarchy approach. These methods will be illustrated in both normal form and multi-stage stochastic game representations and the implications for the role of learning in games will be discussed.
Host: David Wolpert, CCS-3, 665-7914