DEC‐MDP/POMDP
Aurélie Beynier, François Charpillet, Daniel Szer, Abdel‐Illah Mouaddib · 2013
Markov decision processes (MDPs) and partially observable Markov decision processes (DEC-POMDPs) are both mathematical models that have been successfully used to formalize sequential decision-theoretic problems under uncertainty. These models rely on different types of hypotheses that can be classified within: i) each agent has a complete knowledge of the system state; ii) each agent has a partial knowledge of the system state; iii) the agents can communicate; iv) the agents cannot communicate. These hypotheses have led to several formalisms. Among them, this chapter reviews the most well-known ones: MMDP, Decentralized MDPs (DEC-MDPs), Decentralized POMDPs (DEC-POMDPs), MTDP, Dec-MDP-Com, COM-MTDP, ND-POMDP, TI-Dec-MDP, OC-Dec-MDP. It also deals with the complexity of computing optimal solutions for the multiagent decision problems described with these formalisms. DEC-POMDPs and DEC-MDPs extend POMDPs and MDPs to multiagent decentralized control. Controlled Vocabulary Terms multi-agent systems; uncertainty handling