The field of reinforcement learning and adaptive control is rapidly advancing, with a focus on developing more robust and efficient algorithms for real-world applications. Recent research has highlighted the importance of addressing the distribution shift problem in transportation networks, with approaches such as meta reinforcement learning and domain randomization showing promise. Another key area of research is the development of more efficient and data-driven methods for adaptive control, such as symbolic dynamics with residual learning. Noteworthy papers in this area include Sym2Real, which demonstrates robust control of quadrotors and racecars using a fully data-driven framework, and SPiDR, which presents a scalable algorithm for safe sim-to-real transfer with provable guarantees. Additionally, researchers are exploring new approaches to reinforcement learning, such as frictional Q-learning and pure exploration via Frank-Wolfe self-play, which show potential for improving performance and efficiency in complex tasks.
Advances in Reinforcement Learning and Adaptive Control
Sources
The Distribution Shift Problem in Transportation Networks using Reinforcement Learning and AI
Sym2Real: Symbolic Dynamics with Residual Learning for Data-Efficient Adaptive Control
Nonconvex Regularization for Feature Selection in Reinforcement Learning
Uncertainty-Based Smooth Policy Regularisation for Reinforcement Learning with Few Demonstrations
SPiDR: A Simple Approach for Zero-Shot Safety in Sim-to-Real Transfer
Central Limit Theorems for Asynchronous Averaged Q-Learning
Asymptotically Optimal Problem-Dependent Bandit Policies for Transfer Learning
Evaluation-Aware Reinforcement Learning
Efficient $\varepsilon$-approximate minimum-entropy couplings
Frictional Q-Learning
Pure Exploration via Frank-Wolfe Self-Play