Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Learning in zero–sum team Markov games using factored value functions

Lagoudakis Michael, Parr, R.

Full record

---

Summary

We present a new method for learning good strategies in zero-sumMarkov games in which each side is composed of multiple agents collaboratingagainst an opposing team of agents. Our method requires fullobservability and communication during learning, but the learned policiescan be executed in a distributed manner. The value function is representedas a factored linear architecture and its structure determines thenecessary computational resources and communication bandwidth. Thisapproach permits a tradeoff between simple representations with little orno communication between agents and complex, computationally intensiverepresentations with extensive coordination between agents. Thus,we provide a principled means of using approximation to combat theexponential blowup in the joint action space of the participants. The approachis demonstrated with an example that shows the efficiency gainsover naive enumeration.

Services