Abstract EN:

A possible way to accelerate reinforcement learning process is to guide the exploration process using prior domain knowledge. This called knowledge transfer, and most transfer algorithms are based on an implicit assumption. They suppose that prior knowledge has a good quality for the current task. If this condition is not true, the learning process will be worse than standard reinforcement learning algorithm (negative transfer). This thesis put forwards some transfer algorithms to avoid this problem, whose can adapts learning process to prior knowledge quality. More precisely, we introduce a parameter called transfer rate, which controls how much prior knowledge will be used. In addition, we propose to optimize the transfer rate in order to make the best use of this policy. Thus, the proposed algorithms provide some robustness, working for all prior knowledge quality level, which was not the case with previous approaches. These algorithms are evaluated in two different problems: a toy problem (the gridworld), and a real complex one (a coach assistant tool). The latter application offers a coach to seize tactical patterns with a graphical interface, and then allows agents to view players doing the same patterns. To meet within a reasonable time, the request of the coach. The reinforcement learning alone is not enough, and transfer our algorithms have been applied to this area with success