Submit

Login Login

Register Register

Home / Journals / CMC / Online First / doi:10.32604/cmc.2026.086535
Journal Menu
Special Issues
Table of Content

All issues

Online First

2026

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

2007

2006

2005

2004

Open Access

ARTICLE

An Improved Safe Soft Actor-Critic Path Planning Algorithm for Autonomous Vehicles Based on a Dual-Stream Q-Network and Dynamic Analytic Hierarchy Process

Shengxuan Dong, Xiongwei Li*
Shijiazhuang Campus, Army Engineering University of PLA, Shijiazhuang, China
* Corresponding Author: Xiongwei Li. Email: email

Computers, Materials & Continua https://doi.org/10.32604/cmc.2026.086535

Received 01 June 2026; Accepted 21 July 2026; Published online 12 August 2026

Abstract

To address the conflict between navigation performance and safety constraints in safe reinforcement learning, this paper proposes Dual Stream-Analytic Hierarchy Process-Safe Soft Actor (DS-AHP-SAC), a safe soft actor-critic algorithm based on a dual-stream Q-network and dynamic Analytic Hierarchy Process (AHP) stratified experience replay. The algorithm achieves a balance between reward maximization and constraint satisfaction through three synergistic designs: (1) decoupling the Q-network into independent navigation and safety value streams to eliminate gradient interference at the Critic level and mitigate gradient competition at the Actor level; (2) constructing a three-criterion dynamic sampling strategy based on AHP, incorporating safety urgency, information value, and scarcity to enable phase-adaptive experience replay; (3) designing a curriculum-scheduling scheme that linearly increases the safety constraint weight during training, preventing policy degradation caused by premature imposition of high safety penalties. Experimental results in a two-dimensional continuous navigation environment demonstrate that DS-AHP-SAC reduces the violation rate by 20.0% compared to unconstrained SAC without sacrificing navigation success rate, while avoiding the policy collapse observed in SAC-Lagrangian. Ablation studies validate the necessity of the dual-stream Q-network and the convergence acceleration effect of dynamic AHP.

Keywords

Safe reinforcement learning; dual-stream Q-network; analytic hierarchy process; experience replay; constrained Markov decision process

  • 137

    View

  • 32

    Download

  • 0

    Like

Related articles
Copyright© 2026 Tech Science Press
© 1997-2026 TSP (Henderson, USA) unless otherwise stated

Share Link