Authors:
Guillaume Debras
1
;
Abdel-Illah Mouaddib
2
;
Laurent Jean Pierre
2
and
Simon Le Gloannec
3
Affiliations:
1
Université de Caen - GREYC, Cordon Electronics DS2i and Airbus Defence and Space, France
;
2
Université de Caen - GREYC, France
;
3
Cordon Electronics DS2i, France
Keyword(s):
Markov Decision Processes, MultiAgent Decision Making, Game Theory and Applications, Robotic Application.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Artificial Intelligence and Decision Support Systems
;
Co-Evolution and Collective Behavior
;
Computational Intelligence
;
Enterprise Information Systems
;
Evolutionary Computing
;
Evolutionary Robotics and Intelligent Agents
;
Game Theory Applications
;
Soft Computing
Abstract:
Multiagent Markov Decision Processes (MMDPs) provide a useful framework for multiagent decision making. Finding solutions to large-scale problems or with a large number of agents however, has been proven to be computationally hard. In this paper, we adapt H-(PO)MDPs to multi-agent settings by proposing a new approach using action groups to decompose an initial MMDP into a set of dependent Sub-MMDPs where each action group is assigned a corresponding Sub-MMDP. Sub-MMDPs are then solved using a parallel Bellman backup to derive local policies which are synchronized by propagating local results and updating the value functions locally and globally to take the dependencies into account. This decomposition allows, for example, specific aggregation for each sub-MMDP, which we adapt by using a novel value function update. Experimental evaluations have been developed and applied to real robotic platforms showing promising results and validating our techniques.