Prise de décision en temps-réel pour des POMDP de grande taille

This paper presents a POMDP approximation method, called RTBSS (Real- Time Be- lief Space Search), which is based on a look-ahead search in order to p lan in a real-time dy- namic environment. The basis of our approach is to avoid computing full po licies in POMDP problems. Our approach is especially motivated by real-time environments where the state space is too large to consider traditional offline algorithms. We then procee d with an online ap- proach to find at each step, the action that maximize the agent expected utility . To this end, we present the formalism behind our approach. Then, we present how the approach was applied on three different environments: Tag, RockSample and the RoboCupResc ue simulation. Let us men- tion finally, that the approach we present was successfully implemented forthe RoboCupRescue 2004 international competition in Lisbon, Portugal where we finished in sec ond position.