The objective of this theme is to explore and rethink theoretical and empirical foundations of reinforcement learning (RL) — and, more broadly, the issues pertaining to decision-making, autonomy, and dynamical systems — by leveraging recent advances in control theory, high-dimensional probability, approximation theory of deep neural nets, and safe RL. Our research will address these shortfalls:
- Most popular empirically successful algorithms are data-inefficient, especially in problems where online exploration is critical.
- The empirical successes are mostly found in simulators (e.g., video/board games). Real-world applications seldom come with high-quality simulators, and there is very little research on performing RL in a way that complies with real-world constraints (e.g., scarcity of data, safety constraints).
- Despite the impressive empirical performance of deep RL and other methods based on universal function approximation, a theory of RL with large state spaces and function approximation is still lacking.