Authors: Guo, Anyuan (2006)Stochastic planning has gained popularity over classical planning in recent years by offering principled ways to model the uncertainty inherent in many real world applications. In particular, Markov decision process (MDP) based approaches are particularly well suited to problems that involve sequential decision making.
Partially observable stochastic games (POSGs) are a multi-agent extension of the MDP framework to decentralized decision making. While elegant and expressive, this model has been shown to be intractable. Many real world problems exhibit additional structure that can be leveraged for computational gain. The focus of this thesis is on a class of problems in which the agents are largely independent but may interact by constraining each other’s actions.
The best-known s...