The concept of reward is fundamental in reinforcement learning with a wide range of applications in natural and social sciences.Seeking an interpretable reward for decision-making that largely shapes the system's ...The concept of reward is fundamental in reinforcement learning with a wide range of applications in natural and social sciences.Seeking an interpretable reward for decision-making that largely shapes the system's behavior has always been a challenge in reinforcement learning.In this work,we explore a discrete-time reward for reinforcement learning in continuous time and action spaces that represent many phenomena captured by applying physical laws.We find that the discrete-time reward leads to the extraction of the unique continuous-time decision law and improved computational efficiency by dropping the integrator operator that appears in classical results with integral rewards.We apply this finding to solve output-feedback design problems in power systems.The results reveal that our approach removes an intermediate stage of identifying dynamical models.Our work suggests that the discrete-time reward is efficient in search of the desired decision law,which provides a computational tool to understand and modify the behavior of large-scale engineering systems using the optimal learned decision.展开更多
基金supported by the Guangdong Basic and Applied Basic Research Foundation(2024A1515011936)the National Natural Science Foundation of China(62320106008)
文摘The concept of reward is fundamental in reinforcement learning with a wide range of applications in natural and social sciences.Seeking an interpretable reward for decision-making that largely shapes the system's behavior has always been a challenge in reinforcement learning.In this work,we explore a discrete-time reward for reinforcement learning in continuous time and action spaces that represent many phenomena captured by applying physical laws.We find that the discrete-time reward leads to the extraction of the unique continuous-time decision law and improved computational efficiency by dropping the integrator operator that appears in classical results with integral rewards.We apply this finding to solve output-feedback design problems in power systems.The results reveal that our approach removes an intermediate stage of identifying dynamical models.Our work suggests that the discrete-time reward is efficient in search of the desired decision law,which provides a computational tool to understand and modify the behavior of large-scale engineering systems using the optimal learned decision.
