Dynamic Programming with Reinforcement Learning for Scheduling Optimization

Abstract

This paper presents a novel approach to scheduling optimization that synergistically combines the strengths of dynamic programming (DP) and reinforcement learning (RL). Traditional DP methods struggle with complex scheduling scenarios due to their exponential computational complexity, particularly when dealing with intricate constraints and a large state space. This work addresses this limitation by employing a hybrid framework where DP generates an initial, feasible schedule and defines a cost function, while an RL agent dynamically refines this schedule based on real-time system state. The RL agent learns a policy to adapt scheduling parameters, effectively pruning the search space and accelerating convergence towards optimal solutions. The core innovation lies in the iterative interaction between DP and RL, leading to a more robust and efficient scheduling process. This approach demonstrates improved scalability and performance compared to pure DP solutions, particularly in scenarios with dynamic and complex constraints. We introduce a framework that balances the deterministic nature of DP with the adaptive capabilities of RL, offering a promising solution for a wide range of scheduling challenges.