Deep Q-Network Based on a New Optimization Method with Applications to Path Planning

ICLR 2026 Conference Submission17353 Authors

Keywords: Deep Q-learning, path planning, optimal control

TL;DR: We employ a new optimization method to enhance the performance of the DQN network.

Abstract: Path planning is an essential part for agents to navigate in complex environments efficiently. Recent advances in conventional methods and learning-based methods have improved adaptability in complex settings. However, balancing computational efficiency, optimality, and safety in different environments remains a critical open problem. In this paper, we propose a novel Q-Learning framework named OCPDQN based on the optimal control method with application to path planning problems. Furthermore, we improve OCPDQN by combining with Gauss-Newton and propose another new framework named GN-OCPDQN to avoid the extensive computation of the Hessian matrix. Compared to traditional deep Q-networks, which rely on the gradient descent method to update network parameters, the proposed methods present a faster convergence rate and higher robustness. The experimental results demonstrate that both OCPDQN and GN-OCPDQN frameworks show better learning performance than existing deep reinforcement learning methods in the path planning task.

Primary Area: optimization

Submission Number: 17353