Risk Aware Reinforcement Learning with Safety Layer

This thesis aims to analyze the effects of combining two different safe reinforcement learning algorithms to cover the shortcomings of each algorithms. Firstly, a safety layer algorithm which corrects the actions leading to error states is implemented. Safety layer is combined with two different risk sensitive reinforcement learning algorithms: a variance constrained deep deterministic policy gradient algorithm and a risk sensitive distributional deep deterministic policy gradients algorithm. The results are evaluated by comparing rewards, episode lengths, action corrections and variance of the returns provided by vanilla deep deterministic policy gradients algorithms and risk-aware deep deterministic policy gradients algorithms combined with safety layers.