Social Model Shaping for Solving Generic DEC-POMDPs

  • Authors:
  • Pradeep Varakantham

  • Affiliations:
  • -

  • Venue:
  • WI-IAT '11 Proceedings of the 2011 IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology - Volume 02
  • Year:
  • 2011

Quantified Score

Hi-index 0.00

Visualization

Abstract

Decentralized Partially Observable Markov Decision Problem, DEC-POMDP is a popular model to representmulti-agent decision making under uncertainty. However, the significant computational complexity involved in solving DECPOMDPshas limited their application. Recently, social model shaping (TREMOR and D-TREMOR algorithms) was introduced as an alternative to solve a sub-class of DEC-POMDPs. While scalability has been improved to even solve hundred agent problems, social model shaping has been restricted to solving a sub-class of DEC-POMDPs called Distributed POMDPs with Coordination Locales (DPCL). To that end, we make two significant contributions: (a) Firstly, we enhance the model shaping approach to solve general DEC-POMDPs where there is no restriction on the agent dependencies, and (b) Secondly, we provide theoretical justification for the model shaping heuristics. The key intuition is that not all interactions between agents occur at every time step. In addition to solving 100 agent problems in weakly coupled domains (due to extension from TREMOR and D-TREMOR), we are able to show that social model shaping achieves comparable performance to leading DEC-POMDP solvers (such as IMBDP, MBDP-OC, PBIP-IPGetc.) on tightly coupled benchmark problems.