Apprentissage par renforcement pour les processus décisionnels de Markov partiellement observés Apprendre une extension sélective du passé

We present a new algorithm that extends the Reinforcement Learning framework to Partially Observed Markov Decision Processes (POMDP). The main idea of our method is to build a state extension, called exhaustive observable, which allow us to define a next processus that is Markovian. We bring the proof that solving this new process, to which classical RL methods can be applied, brings an optimal solution to the original POMDP. We apply the algorithm built on that proof to several examples to test its validity and robustness. MOTS-CLÉS : théorie de la décision, agent intelligent, apprentissage par renforcement, POMDP, extension d’état