Lazy-MDPs: Towards Interpretable Reinforcement Learning by Learning When to Act
This addresses the interpretability issue in reinforcement learning for AI agents, though it is incremental as it builds on standard MDPs with a new mode of action.
The paper tackles the problem of making reinforcement learning agents more interpretable by enabling them to decide when to act, rather than just how, through lazy-MDPs, which allow deferring to a default policy and penalize non-lazy actions, resulting in competitive performance in Atari games with control taken in only a limited subset of states.
Traditionally, Reinforcement Learning (RL) aims at deciding how to act optimally for an artificial agent. We argue that deciding when to act is equally important. As humans, we drift from default, instinctive or memorized behaviors to focused, thought-out behaviors when required by the situation. To enhance RL agents with this aptitude, we propose to augment the standard Markov Decision Process and make a new mode of action available: being lazy, which defers decision-making to a default policy. In addition, we penalize non-lazy actions in order to encourage minimal effort and have agents focus on critical decisions only. We name the resulting formalism lazy-MDPs. We study the theoretical properties of lazy-MDPs, expressing value functions and characterizing optimal solutions. Then we empirically demonstrate that policies learned in lazy-MDPs generally come with a form of interpretability: by construction, they show us the states where the agent takes control over the default policy. We deem those states and corresponding actions important since they explain the difference in performance between the default and the new, lazy policy. With suboptimal policies as default (pretrained or random), we observe that agents are able to get competitive performance in Atari games while only taking control in a limited subset of states.