Policy-based Methods for Realistic Reinforcement Learning Applications
Reinforcement Learning (RL) allows an agent to learn how to behave optimally in sequential decision-making problems.
Among the various paradigms in RL, policy-based algorithms have reached notable results when dealing with real-world continuous control problems.
Policy-based methods learn stochastic (hyper-)policies by exploring the action space or the parameter space, and they present some pitfalls in real-life usage.
For instance, in real-world applications stochastic controllers are undesirable, since they lack of robustness, safety, and traceability.
In this work, we study novel approaches within the policy-based class of methods to address some of the problem arising when applying RL to real-world scenarios, such as deploying a deterministic controllers, or satisfying some structural and user-defined constraints.
Back to Current Students