|CANONACO GIUSEPPE||Cycle: XXXIV |
Section: Computer Science and Engineering
Tutor: GATTI NICOLA
Advisor: ROVERI MANUEL Major Research topic
:Non-stationary Reinforcement LearningAbstract:
In most reinforcement learning studies the considered task is often assumed to be stationary, i.e., it does not change its characteristics over time. Unfortunately, this assumption does not hold in real-world scenarios, where, usually, systems and environments evolve with time. By evolving with time we mean that at some fixed point in time, a-priori not known, the markov decision process modeling our reinforcement learning task changes either in terms of reward function or state-transition function (possibly both of these functions can be affected by a non-stationary behavior). Since the convergence properties of standard reinforcement learning algorithms are strongly based on stationarity assumptions, in this new context they will not work appropriately. Therefore, the main objective of this project consist of devising reinforcement learning algorithms able to correctly deal with non-stationary environments and systems.