Towards Reliable Deep Reinforcement Learning for Industrial Applications: A DDPG-based Algorithm with Improved Performance

Abstract

This paper proposes an Improved Model-Based Deep Deterministic Policy Gradient, a novel reinforcement learning algorithm designed to overcome three critical challenges in industrial deep reinforcement learning applications: (1) poor sample efficiency requiring excessive real-world trials, (2) safety risks from unstable policies during training, and (3) difficulty scaling to high-dimensional continuous control spaces. Building on DDPG’s strengths for continuous control, the proposed algorithm introduces four key innovations: (i) a virtual environment for data-efficient learning, (ii) a simulation rate mechanism adapting model reliance dynamically, (iii) a simulated experience buffer preventing divergence, and (iv) a performance threshold for fail-safe operation. Evaluated on the Cart-Pole benchmark via the OpenAI Gym Python library, the suggested method demonstrates faster convergence than standard DDPG while maintaining performance degradation under sensor malfunctions or communication losses. These improvements derive from the algorithm’s unique ability to simultaneously leverage real-world data and model-generated experiences, reducing physical trial costs while ensuring operational safety. The results establish the novel framework as a practical solution for industrial control systems where reliability and data efficiency are paramount, particularly in applications like chemical process control and precision robotics that demand stable operation amid sensor/communication failures. © 2026, Amirkabir University of Technology. All rights reserved.