One-Step Distributional Reinforcement Learning
This work addresses theoretical limitations in DistrRL for researchers, offering a more tractable framework with proven convergence, though it is incremental as it builds on existing DistrRL methods.
The paper tackles the theoretical gaps in distributional reinforcement learning (DistrRL) by proposing a simpler one-step framework (OS-DistrRL) that focuses on randomness from one-step dynamics, and it shows that this approach provides unified convergence theory and outperforms categorical DistrRL in experiments.
Reinforcement learning (RL) allows an agent interacting sequentially with an environment to maximize its long-term expected return. In the distributional RL (DistrRL) paradigm, the agent goes beyond the limit of the expected value, to capture the underlying probability distribution of the return across all time steps. The set of DistrRL algorithms has led to improved empirical performance. Nevertheless, the theory of DistrRL is still not fully understood, especially in the control case. In this paper, we present the simpler one-step distributional reinforcement learning (OS-DistrRL) framework encompassing only the randomness induced by the one-step dynamics of the environment. Contrary to DistrRL, we show that our approach comes with a unified theory for both policy evaluation and control. Indeed, we propose two OS-DistrRL algorithms for which we provide an almost sure convergence analysis. The proposed approach compares favorably with categorical DistrRL on various environments.