|Reinforcement learning (RL) can optimally solve decision and control problems involving complex dynamic systems, without requiring a mathematical model of the system. Online RL algorithms do not even require data in advance; they learn from experience. If a model is available, dynamic programming (DP), the model-based counterpart of RL, can be used. RL and DP are applicable in a variety of disciplines, including automatic control, artificial intelligence, economics, and medicine. However, to scale up to realistic control problems, RL and DP must employ compact, approximate representations of the solution.
In this project, we are investigating sound and efficient algorithms for approximate RL and DP, focusing on control applications. We study the convergence and performance properties of these algorithms, and evaluate them in comprehensive benchmarks on a range of nonlinear control problems. In collaboration to MSc and BSc students, we investigate a number of practical issues in RL, such as applications to real-time robotic control and methods to accelerate learning.