Accelerating DQN with Transfer Learning for Industrial Process Scheduling under Environmental Changes

In industrial process optimization field, solution methods could fail due to minor environmental changes. Re-optimization starting from scratch under new environment could be costly, especially using reinforcement learning. To address this, we propose a Deep Q-Network (DQN)-based transfer learning acceleration method (DTLA) that incorporates probability distributions and pseudo-labels to quickly obtain satisfactory solutions. The method filters the source domain experience using frequency statistics and refines the selection by incorporating rewards from the target domain, generating a pseudo-label generation function. This function creates pseudo-labels for the source domain experience, which are used as initial experience for the target domain, accelerating the solution process. Experimental results show that the proposed method yields feasible solutions faster than genetic algorithms and network parameter transfer methods, with strong performance.