Current students


CESTERO PORTU JULENCycle: XXXVI

Section: Computer Science and Engineering
Advisor: RESTELLI MARCELLO
Tutor: GATTI NICOLA

Major Research topic:
Bridging the gap between Reinforcement Learning and Planning

Abstract:
Reinforcement Learning has been proven to be a very adaptable solution to general problems. On the other hand, planning algorithms provide good results for specific situations. Although Reinforcement Learning learns a policy that can be used in any state, it can be ineffective for large state spaces and infrequent states. Planning algorithms, on the other hand, can accurately estimate the utility of the different actions in the current state, even for large state spaces, but they have the drawback that the planning time for each step is very long. In this thesis, we propose a hybrid approach combining the speed and adaptability of Reinforcement Learning with the accuracy of planning algorithms, which counters the disadvantages of both types of algorithms used separately. This hybrid approach is validated in a real logistics case where an agent trained with a mix of Reinforcement Learning and planning algorithms must manage the inventory of a storehouse. The results are validated on real-world industrial datasets.