Approaching Single-Episode Survival Reinforcement Learning with Safety-Threshold Q-Learning
Résumé
Survival Reinforcement Learning is a specific type of RL problem constrained by a risk of ruin. The underlying stochastic sequential decision process with which the agent interacts includes a budget that evolves over time with the received rewards and must remain positive throughout its entire lifetime. The goal is to find a good trade-off between exploration, exploitation, and safety during a single learning episode, maximizing rewards while managing the available budget to minimize the probability of ruin. Existing approaches do not provide satisfactory solutions to this problem. This paper introduces the safety-threshold heuristic, which is used to extend the standard Q-Learning method. A simulated grid environment is used to evaluate its performance.
Origine | Fichiers produits par l'(les) auteur(s) |
---|