<p>Due to the everchanging dynamics of traffic situation, managing real-time traffic congestion with great efficiency is exceedingly challenging. Deep Reinforcement Learning (DRL) in Intelligent Transportation System (ITS) under the concept of Edge computing is an approach that determines the optimal traffic signal strategy for dealing with traffic congestion. Optimizing traffic signal with a DRL agent involves transmitting state information collected by edge devices. However, network congestion, device malfunctions, and transmission delays often impede the transmission of information. Consequently, the decision-making capacity of the agent suffers from inadequate information, leading to decreased efficacy. To mitigate this issue, the study proposes two distinct masking methods on input states. A single DRL agent deals with these masked inputs from the environment through the Edge devices. In order to train the agent, the DRL algorithm Proximal Policy Optimization (PPO) is implemented in five different neural network models including the state-of-the-art Transformer network which can accurately model spatial dependence and capture the persistence of sequential data. To validate the feasibility of the agent, simulation experiments are conducted in hypothetical road network and real-time road map. The experiments utilize waiting time, fuel consumption, and <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="13177_2025_482_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="34" /> </InlineMediaObject> <EquationSource Format="TEX">$CO_{2}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>C</mi> <msub> <mi>O</mi> <mn>2</mn> </msub> </mrow> </math></EquationSource> </InlineEquation> emission as key simulation metrics due to their significant impact on traffic congestion. However, the main goal is to alleviate traffic congestion by minimizing waiting time. Results demonstrate substantial reductions in waiting times for both networks, with reductions of 26.35% and 26.31% observed for the two masking strategies in the hypothetical scenario, and decreases of 5.86% and 6.86% recorded for the real-time road map, highlighting significant improvements in congestion alleviation efforts.</p>

DRL Based Traffic Signal Control Method Featuring Masked Approach to Redress Transmission Error in ITS

期刊论文
发表日期：

2025年4月2日

查看全文

Ananya Paul,
Isha Ganguli,
Rajat Subhra Bhowmick,
Sumit Badotra,
Salil Bharany,
Ateeq Ur Rehman

摘要

Due to the everchanging dynamics of traffic situation, managing real-time traffic congestion with great efficiency is exceedingly challenging. Deep Reinforcement Learning (DRL) in Intelligent Transportation System (ITS) under the concept of Edge computing is an approach that determines the optimal traffic signal strategy for dealing with traffic congestion. Optimizing traffic signal with a DRL agent involves transmitting state information collected by edge devices. However, network congestion, device malfunctions, and transmission delays often impede the transmission of information. Consequently, the decision-making capacity of the agent suffers from inadequate information, leading to decreased efficacy. To mitigate this issue, the study proposes two distinct masking methods on input states. A single DRL agent deals with these masked inputs from the environment through the Edge devices. In order to train the agent, the DRL algorithm Proximal Policy Optimization (PPO) is implemented in five different neural network models including the state-of-the-art Transformer network which can accurately model spatial dependence and capture the persistence of sequential data. To validate the feasibility of the agent, simulation experiments are conducted in hypothetical road network and real-time road map. The experiments utilize waiting time, fuel consumption, and $CO_{2}$ $C O_{2}$ emission as key simulation metrics due to their significant impact on traffic congestion. However, the main goal is to alleviate traffic congestion by minimizing waiting time. Results demonstrate substantial reductions in waiting times for both networks, with reductions of 26.35% and 26.31% observed for the two masking strategies in the hypothetical scenario, and decreases of 5.86% and 6.86% recorded for the real-time road map, highlighting significant improvements in congestion alleviation efforts.