CEM: Constrained Entropy Maximization for Task-Agnostic Safe Exploration

In the absence of assigned tasks, a learning agent typically seeks to explore its environment efficiently. However, pursuit exploration will bring more safety risks. An under-explored aspect reinforcement is how achieve safe efficient when task unknown. this paper, we propose practical Constrained Entropy Maximization (CEM) algorithm solve task-agnostic problems, which naturally require finite horizon and undiscounted constraints on costs. The CEM aims learn policy that maximizes state entropy under premise safety. To avoid approximating density in complex domains, leverages k-nearest neighbor estimator evaluate efficiency exploration. terms safety, minimizes costs, adaptively trades off based current constraint satisfaction. empirical analysis shows enables acquisition environments, resulting improved performance both sample for target tasks.