Exploring Safer Behaviors for Deep Reinforcement Learning

We consider Reinforcement Learning (RL) problems where an agent attempts to maximize a reward signal while minimizing cost function that models unsafe behaviors. Such formalization is addressed in the literature using constrained optimization on cost, limiting exploration and leading significant trade-off between reward. In contrast, we propose Safety-Oriented Search complements Deep RL algorithms bias policy toward safety within evolutionary optimization. leverage benefits design novel concept of safe mutations use visited states explore safer actions. further characterize behaviors policies over desired specifics with sample-based bound estimation, which makes prior verification analysis tractable training loop. Hence, driving learning process towards regions space. Empirical evidence Safety Gym benchmark shows successfully avoid drawbacks return improving policy.