Boundary-aware Reinforcement Learning for Hypercube State Spaces via Deterministic Policy Gradient
The authors developed a continuous-time deep policy gradient algorithm for learning control policies in reflected stochastic systems on a hypercube. They linked the value function to Neumann Bellman equations to reduce errors at boundaries, leading to a stable algorithm called CT-DDPG-R. This method uses boundary-aware constraints—either soft penalties or hard architectural features—to significantly improve learning stability and final performance compared to standard methods.