Deep Reinforcement Learning-Based Driving Strategy for Avoidance of Chain Collisions and Its Safety Efficiency Analysis in Autonomous Vehicles

Muzahid, Abu Jafar Md; Kamarulzaman, Syafiq Fauzi; Rahman, Md. Arafatur; Alenezi, Ali H. · 2022 · DOAJ

DOI: 10.1109/ACCESS.2022.3167812

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical challenge of chain collision avoidance in autonomous vehicle (AV) traffic flows, a problem exacerbated by unexpected critical situations such as sudden decelerations and abrupt lane changes. While existing methods often rely on static obstacle avoidance or high-computation prediction models that struggle with dynamic environments, this study proposes a Deep Reinforcement Learning (DRL) based driving strategy. The research is motivated by the need to improve safety efficiency and reliability in multi-agent AV systems, where traditional reaction-based or prediction-based techniques fail to account for future states or kinematic limitations effectively. The methodology frames chain collision avoidance as a Markov Decision Process (MDP). The authors developed a perception network structure utilizing actor-critic methodologies, specifically employing Long Short-Term Memory (LSTM) networks to handle temporal features and sensor inputs like camera and 2D-lidar data. The state space includes vehicle positions and speeds within a 20-meter range, while the action space comprises lateral maneuvers (stay, change left/right) and longitudinal controls (acceleration, braking). A custom reward function was designed to balance driving behavior and collision avoidance. The strategy was evaluated using three state-of-the-art actor-critic algorithms in Unity3D simulations, covering both single-agent and multi-agent environments. The analysis focused on three traffic scenarios: sudden slowdowns, abrupt lane changes, and smooth destination reaching, assessing metrics such as training performance, success rate, and reward stability. The findings demonstrate that the proposed DRL-based strategy effectively mitigates chain collisions by learning optimal driving policies through interaction with the environment. The study highlights the advantages of using LSTM for its lower computational cost and ability to extract temporal features without requiring massive datasets. The comparative analysis of the three actor-critic algorithms revealed differences in training speed, stability, and the trade-off between exploration and exploitation. The results indicate that the agent can successfully navigate complex, uncertain traffic flows, maintaining safety and comfort while avoiding secondary incidents caused by cumulative reaction times or narrow vehicle-to-vehicle distances. The significance of this work lies in its contribution to the development of robust, reliable autonomous traffic systems. By providing a comprehensive safety efficiency analysis, the paper offers insights for academics and policymakers on optimizing AV control logic. The proposed method reduces the reliance on precise mathematical models and high-quality maps, making it adaptable to real-world uncertainties. Ultimately, the study paves the way for the practical deployment of driver-less traffic systems by addressing the specific weaknesses of existing DRL algorithms in multi-vehicle collision avoidance scenarios.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• secondary collisions