Traffic state adversarial disturbance generation method for single intersection signal control based on fast gradient descent

Abstract

A traffic state confrontation disturbance generation method based on fast gradient descent single intersection signal control, based on the existing reinforcement learning DQN algorithm training traffic intersection signal light control model, using FGSM attack and combining the gradient value to discretize the confrontation disturbance The adversarial samples are generated by the optimization process, and the final perturbation state is input into the agent model by combining the adversarial disturbance with the original state. Finally, the smoothness or congestion degree of the single intersection is tested on sumo. The present invention can limit the size of the disturbance under the condition that the output disturbance has physical meaning, thereby efficiently generating the confrontation state, increasing the queuing length and waiting time at the intersection, greatly reducing the performance of the model, and greatly reducing the flow rate of the traffic intersection.

Description

technical field [0001] The invention belongs to the intersection field of intelligent traffic and machine learning information security, and relates to a traffic state confrontation disturbance generation method based on fast gradient descent (FGSM) single-intersection signal control. Background technique [0002] Traffic congestion has become an urgent challenge for urban transportation, and one of the most critical considerations when designing a modern city is the development of an intelligent traffic management system. The main goal of a traffic management system is to reduce traffic congestion, which is one of the major problems in big cities these days. Efficient urban traffic management saves time and money, and reduces carbon dioxide emissions into the atmosphere. [0003] Reinforcement learning (RL) has yielded impressive results as a machine learning technique for traffic signal control problems. Reinforcement learning does not require prior comprehensive knowled...

AI Technical Summary

Problems solved by technology

Although DRL has great advantages, it is vulnerable to adversarial attacks, such as: luring attacks, policy timing attacks, value function-based adversarial attacks, Trojan horse attacks, etc.

Images