On the Theory of Reinforcement Learning with Once-per-Episode Feedback
This addresses a more realistic feedback scenario for applications like self-driving cars and robotics, though it is a theoretical incremental contribution.
The paper tackles reinforcement learning with binary feedback only at episode end, showing learning is possible via an algorithm achieving sublinear regret under an unknown parametric model.
We study a theory of reinforcement learning (RL) in which the learner receives binary feedback only once at the end of an episode. While this is an extreme test case for theory, it is also arguably more representative of real-world applications than the traditional requirement in RL practice that the learner receive feedback at every time step. Indeed, in many real-world applications of reinforcement learning, such as self-driving cars and robotics, it is easier to evaluate whether a learner's complete trajectory was either "good" or "bad," but harder to provide a reward signal at each step. To show that learning is possible in this more challenging setting, we study the case where trajectory labels are generated by an unknown parametric model, and provide a statistically and computationally efficient algorithm that achieves sublinear regret.