<ul><li>Constrained Reinforcement Learning (CRL) focuses on sequential decision-making with goal achievement while meeting constraints.</li><li>Policy-based methods, especially in continuous-control problems, feature action-based or parameter-based exploration strategies.</li><li>A new exploration-agnostic algorithm, C-PG, is introduced with global convergence guarantees under gradient domination assumptions.</li><li>C-PG demonstrates effectiveness in learning deterministic policies for constrained control tasks based on empirical validation and comparisons with baselines.</li></ul>

Learning Deterministic Policies with Policy Gradients in Constrained Markov Decision Processes

Discover more