FLight: Optimization Algorithm for Traffic Lights Based on Short-Term Traffic State Forecast

Abstract

Intelligent transportation systems play a crucial role in building smart cities. However, existing traffic signal control systems heavily rely on knowledge of experts and lack real-time adaptability to dynamic traffic states. Most reinforcement learning-based approaches define traffic states using features such as current signal phases and queue lengths in each lane. Unfortunately, these methods overlook the spatiotemporal characteristics of real-world traffic flow and also fail to consider the moving vehicles when designing signal strategies. In this paper, we propose a Short-Term Traffic Flow Forecast Model that captures both temporal and spatial features of traffic flow. Our approach combines Temporal Convolutional Networks for extracting time series features from historical traffic flow, and Long Short-Term Memory networks to model complex temporal dependencies, and obtains more accurate time dependency relationships by assigning weights through self-attention layers. We also use Graph Attention Networks to learn spatial features between intersections and neighboring junctions. In addition, we have proposed an Enhanced Traffic State representation that captures both waiting and moving vehicles. We divide the lane into multiple segments based on the distance to the intersection and count the number of vehicles driving on different segments. In comparison to conventional traffic signal control methods and several reinforcement learning-based algorithms, our proposed algorithm has achieved superior results, especially, the performance can be increased by up to 4.57% on the real-world datasets.