Multi Task Inverse Reinforcement Learning for Common Sense Reward
This addresses the problem of reward hacking for AI agents in complex environments, though it appears incremental as it builds on existing inverse reinforcement learning methods.
The paper tackles the challenge of reward misalignment in reinforcement learning by proposing to disentangle rewards into task-specific and common-sense components, showing that multi-task inverse reinforcement learning can learn a useful reward function from expert demonstrations.
One of the challenges in applying reinforcement learning in a complex real-world environment lies in providing the agent with a sufficiently detailed reward function. Any misalignment between the reward and the desired behavior can result in unwanted outcomes. This may lead to issues like "reward hacking" where the agent maximizes rewards by unintended behavior. In this work, we propose to disentangle the reward into two distinct parts. A simple task-specific reward, outlining the particulars of the task at hand, and an unknown common-sense reward, indicating the expected behavior of the agent within the environment. We then explore how this common-sense reward can be learned from expert demonstrations. We first show that inverse reinforcement learning, even when it succeeds in training an agent, does not learn a useful reward function. That is, training a new agent with the learned reward does not impair the desired behaviors. We then demonstrate that this problem can be solved by training simultaneously on multiple tasks. That is, multi-task inverse reinforcement learning can be applied to learn a useful reward function.