Learning to Act in Decentralized Partially Observable MDPs

Abstract : We address a long-standing open problem of reinforcement learning in decentralized partially observable Markov decision processes. Previous attempts focussed on different forms of generalized policy iteration, which at best led to local optima. In this paper, we restrict attention to plans, which are simpler to store and update than policies. We derive, under certain conditions, the first near-optimal cooperative multi-agent reinforcement learning algorithm. To achieve significant scalability gains, we replace the greedy maximization by mixed-integer linear programming. Experiments show our approach can learn to act near-optimally in many finite domains from the literature.
Complete list of metadatas

Cited literature [48 references]  Display  Hide  Download

https://hal.inria.fr/hal-01851806
Contributor : Jilles Steeve Dibangoye <>
Submitted on : Monday, July 30, 2018 - 11:51:21 PM
Last modification on : Tuesday, November 19, 2019 - 12:12:55 PM
Long-term archiving on : Wednesday, October 31, 2018 - 2:02:29 PM

File

paper.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-01851806, version 1

Citation

Jilles Steeve Dibangoye, Olivier Buffet. Learning to Act in Decentralized Partially Observable MDPs. ICML 2018 - 35th International Conference on Machine Learning, Jul 2018, Stockholm, Sweden. pp.1233-1242. ⟨hal-01851806⟩

Share

Metrics

Record views

267

Files downloads

77