Inverse Reinforcement Learning from Failure
Kyriacos Shiarlis‚ Joao Messias‚ Maarten van Someren and Shimon Whiteson
In this paper, we approach the problem of Inverse Reinforcement Learning (IRL) from a rather different perspective. Instead of trying to only mimic an expert as in traditional IRL, we present a method that can also utilise failed or bad demonstrations of a task. In particular, we propose a new IRL algorithm that extends the state-of-the-art method of Maximum Causal Entropy Inverse Reinforcement Learning to exploit such failed demonstrations. Furthermore, we present experimental results showing that our method can learn faster and better than its original counterpart.