Current students

In Reinforcement Learning (RL) an agent acts in an unknown, or not completely known, environment with the goal of maximizing an external reward signal.
We focus our study on the case in which multiple simulators (i.e. fidelity) of the environment are given to the agent. Each of these representations of the world is related to a precision-cost pair with the following meaning: a simulator produces data of a given quality (i.e., precision) but it requires a certain cost (e.g., time, energy).
The key question that we aim to answer is how to exploit this additional possibility to minimize the cost of the training procedure.