Explainable Reinforcement Learning with Causality

The project's principal goal is to improve the safety and explainability of reinforcement learning (RL) systems by utilising causal methods.

We intend to employ causal techniques to decipher the reasons behind an RL agent's specific actions and global policies. We also may investigate which experiences are pivotal for an agent's learning process and to identify the elements of the agent's experiences that explain its optimal policy.

To meet these objectives, our plan is to:
1. Examine the current state-of-the-art methods in explainable AI, with a particular focus on explainable reinforcement learning (XRL).
2. Investigate and select the most pertinent causal techniques for XRL.
3. Apply causal graph discovery methods to elucidate the relationships and causal impact of system variables (like feature representations of states) on the agent's decisions or policies.
4. Evaluate our results against other existing methods in XRL in terms of faithfulness of generated explanations to agent's behaviour and explanation quality.

To date, there is only a single paper that integrates causal methods within XRL (https://arxiv.org/abs/1905.10958), which is limited to explaining individual actions without the capacity to explain policies. Moreover, no existing XRL methodologies incorporate the causal graph discovery, which is an important part of our research.

This doctoral research is relevant to the following research area under the Engineering and Physical Sciences Research Council (EPSRC): Artificial Intelligence Technologies.