Towards possibilistic reinforcement learning algorithms

We propose a framework and algorithms for reinforcement learning in sequential decision problems under uncertainty in which the rewards are qualitative, and/or am temporarily aggregated by a "minimum" instead of a sum as in the classical Markov Decision Processes (MDP) framework. The framework is based on a "possibilistic" version of Markov Decision Processes and the learning algorithms are based on indirect methods in which the possibilistic model of the problem is learned while the problem itself is solved, using Dynamic Prong.