Affinity-Based Reinforcement Learning : A New Paradigm for Agent Interpretability

Abstract

The steady increase in complexity of reinforcement learning (RL) algorithms is accompanied by a corresponding increase in opacity that obfuscates insights into their devised strategies. Methods in explainable artificial intelligence seek to mitigate this opacity by either creating transparent algorithms or extracting explanations post hoc. A third category exists that allows the developer to affect what agents learn: constrained RL has been used in safety-critical applications and prohibits agents from visiting certain states; preference-based RL agents have been used in robotics applications and learn state-action preferences instead of traditional reward functions. We propose a new affinity-based RL paradigm in which agents learn strategies that are partially decoupled from reward functions. Unlike entropy regularisation, we regularise the objective function with a distinct action distribution that represents a desired behaviour; we encourage the agent to act according to a prior while learning to maximise rewards. The result is an inherently interpretable agent that solves problems with an intrinsic affinity for certain actions. We demonstrate the utility of our method in a financial application: we learn continuous time-variant compositions of prototypical policies, each interpretable by its action affinities, that are globally interpretable according to customers’ financial personalities. Our method combines advantages from both constrained RL and preferencebased RL: it retains the reward function but generalises the policy to match a defined behaviour, thus avoiding problems such as reward shaping and hacking. Unlike Boolean task composition, our method is a fuzzy superposition of different prototypical strategies to arrive at a more complex, yet interpretable, strategy.