Planning under Uncertainty in Constrained Multiagent Systems

14^th November 2017, 13:00 Ashton Lecture Theater
Dr. Matthijs Spaan
Algorithmics group,
Delft University of Technology

Abstract

Developing algorithms for multiagent planning under uncertainty is an important area of Artificial Intelligence given that intelligent distributed systems are becoming ubiquitous in society. In particular, when decision-making agents share collectively owned resources, their actions need to be coordinated subject to the availability of these resources.

In this talk, I will discuss recent work on constrained multiagent sequential decision making, focusing on scenarios in which agents need to satisfy resource constraints in expectation but cannot coordinate during execution. However, by bounding expected resource consumption there is no guarantee on the probability of actual constraint violations. In our work, we use Hoeffding's inequality to compute reduced resource limits such that the violation probability is bounded. As the resulting solution is often overly conservative, we relax it dynamically to achieve maximum value subject to a user-defined tolerance of resource violations.

I will conclude by discussing several extensions, such as dealing with partial observability or stochastic resource constraints.