Reinforcement learning (RL) is a subfield of machine learning aimed at solving sequential decision-making problems. RL learns an optimal way of behaving, called policy, by interacting with an environment. Classical methods rely solely on real experience collected through direct interaction with the environment. However, constantly querying the environment can be costly in practical applications, where minimizing the number of such interactions is crucial.
One way to address this issue is to learn, alongside the policy, a model of the environment that can be used in place of the environment itself to gather information about the outcomes of selected actions. In this seminar, we revisit the main ideas of model-based RL, discussing its improvements over model-free RL as well as its potential limitations.

The seminar can be also be followed online on Webex.