PATEL, Anil Ranjitbhai (RPTU Kaiserslautern), Shah, Tirth Maheshkumar (RPTU Kaiserslautern), Liggesmeyer, Peter (RPTU Kaiserslautern)

Adaptive Risk Feature Thresholds in Automated Driving Systems: A Deep Q-Learning Approach

Scheduled for presentation during the Regular Session "Driving based on reinforcement learning" (ThAT6), Thursday, September 26, 2024, 11:50−12:10, Salon 14

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on October 8, 2024

Abstract

In the domain of Automated Driving Systems (ADS), Risk Features (RFs) are pivotal in assessing the risk associated with the vehicle's operational environment. However, the mathematical equations for these RFs often lack sufficient context information, making them less effective in evaluating the system's overall risk.

To tackle this challenge, we introduce a learning-based Deep Q-Network (DQN) model that adapts to dynamic environments and iteratively refines its thresholds in response to various driving conditions. The adaptability of the DQN model, along with its proficiency in handling high-dimensional state spaces, makes it particularly suitable for complex scenarios where traditional RF equations might fall short. This approach ensures more precise and context-aware RF thresholds for ADS, taking into account environmental factors such as heavy rain and changes in vehicle mass, as well as system-specific characteristics.

Our research reveals that traditional mathematical approaches to RF calculation, primarily based on kinematic relations such as distances and velocities, often overlook the broader context by failing to consider environmental and system capabilities. In contrast, our DQN model effectively predicts optimized, context-aware adaptive thresholds that account for both the system's capabilities and the surrounding environment. This method offers an effective approach to establishing RF thresholds in ADS.