Robust Reinforcement Learning in Continuous Action/State Space

Abstract: In this project we aim to apply Robust Reinforce-ment Learning algorithms, presented by Doya and Morimoto [1],[2], to control problems. Specifically, we train an agent to balancea pendulum in the unstable equilibrium, which is the invertedstate.We investigate the performance of controllers based on twodifferent function approximators. One is quadratic, and the othermakes use of a Radial Basis Function neural network. To achieverobustness we will make use of an approach similar toH∞control, which amounts to introducing an adversary in the controlsystem.By changing the mass of the pendulum after training, we aimedto show as in [2] that the supposedly robust controllers couldhandle this disruption better than its non-robust counterparts.This was not the case. We also added a random disturber signalafter training and performed similar tests, but we were againunable to show robustness.