Pitfalls of learning a reward function online
This work addresses foundational challenges in reinforcement learning and agent design, particularly for systems like inverse reinforcement learning, by highlighting and mitigating risks in online reward learning, which is incremental in formalizing desirable properties.
The paper investigates the pitfalls of learning a reward function online in a continual learning setting, where an agent simultaneously learns and optimizes its reward function, leading to issues like manipulation, refusal to learn, and dominated decisions. It introduces formal properties of unriggability and uninfluenceability to address these problems, proving their equivalence under certain conditions.
In some agent designs like inverse reinforcement learning an agent needs to learn its own reward function. Learning the reward function and optimising for it are typically two different processes, usually performed at different stages. We consider a continual (``one life'') learning approach where the agent both learns the reward function and optimises for it at the same time. We show that this comes with a number of pitfalls, such as deliberately manipulating the learning process in one direction, refusing to learn, ``learning'' facts already known to the agent, and making decisions that are strictly dominated (for all relevant reward functions). We formally introduce two desirable properties: the first is `unriggability', which prevents the agent from steering the learning process in the direction of a reward function that is easier to optimise. The second is `uninfluenceability', whereby the reward-function learning process operates by learning facts about the environment. We show that an uninfluenceable process is automatically unriggable, and if the set of possible environments is sufficiently rich, the converse is true too.