Alessandro Lazaric is giving an invited talk at the “Journée pour la promotion et le developpement de l’intelligence artificielle” organized by AfIA.
Reinforcement Learning and Transfer of Knowledge
Reinforcement learning (RL) studies the problem of learning how to behave optimally in uncertain environments through direct experience. While many techniques are by now available to learn near-optimal behaviors, one of their main limitations is that, once a problem is solved, the RL process has to be restarted from scratch every time the task changes. This limitation can be overcome by extracting general knowledge from the solution of a task, and transfer it when solving novel (related) tasks. In this talk, the RL framework will be introduced together with its main challenges and I will provide examples on how knowledge can be extracted and transferred and how this can improve the learning performance across tasks.