Traffic light control method based on heuristic deep Q network

Abstract

The invention discloses a multi-intersection traffic signal control method based on heuristic deep Q learning, and mainly solves problems of correlation of training data, incapability of rapid convergence of a traffic light control strategy and low control efficiency in an existing method. The method comprises steps that the urban traffic network information is read, a vehicle traffic state set ofeach intersection is established, and the read urban traffic network information is converted into an adjacent matrix to be stored; a state set, an action set and an action reward value of each intersection are acquired from a vehicle traffic state set of each intersection; according to the state set, the action set, the action reward value and the adjacency matrix, a heuristic deep Q network method is used for continuously executing actions according to the state of each intersection to obtain rewards and then to the next state, and urban road network traffic lights are controlled. The method can improve control efficiency of intersection traffic signal lamps, improves performance of the multi-intersection traffic signal controller, can be used for urban traffic management, and reduces the urban traffic congestion.

Description

technical field [0001] The invention belongs to the field of intelligent traffic control, in particular to a traffic light control method, which can be used for urban traffic management and reduce urban traffic congestion. Background technique [0002] In the field of urban intelligent traffic control, deep learning and reinforcement learning are currently very popular research directions and have achieved good results. Reinforcement learning is to learn the optimal traffic control strategy by continuously interacting with the urban road environment to obtain the environmental state, and form an adaptive control system for urban traffic. However, due to the increase in the complexity of the urban road environment, the dimension of the state-action space will increase dramatically in the process of acquiring prior knowledge. In order to solve such problems, the deep reinforcement learning DRL formed by the combination of reinforcement learning and deep learning uses both the...

AI Technical Summary

Problems solved by technology

However, in this model, each agent learns independently and does not cooperate with each other, which makes the convergence speed of the deep reinforcement learning algorithm very slow, and it is difficult to learn an optimal traffic light control strategy.

Images