<p>This research aims to find a schedule that adapts to the variations in the transport system that influence the traveled time by the buses. For this purpose, the stop-skipping control strategy is adopted with the goal to serve all passengers waiting at stations with a minimum total delay in a well known static environment. We introduce a novel measure to calculate the maximum total delay of passengers based on the notion of balancing the load inside the buses, denoted as <InlineEquation ID="IEq1"> <EquationSource Format="TEX">${\text { load-delay }}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mspace width="0.333333em" /> <mtext>load-delay</mtext> <mspace width="0.333333em" /> </mrow> </math></EquationSource> </InlineEquation>. To solve the stop-skipping decision problem with minimizing the <InlineEquation ID="IEq2"> <EquationSource Format="TEX">${\text { load-delay }}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mspace width="0.333333em" /> <mtext>load-delay</mtext> <mspace width="0.333333em" /> </mrow> </math></EquationSource> </InlineEquation>, we model it as a distributed game where the stations are the players. We explore three stateless reinforcement learning algorithms: Linear Reward Inaction (LRI), Upper Confidence Bound (UCB), and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$\epsilon $</EquationSource> <EquationSource Format="MATHML"><math> <mi>ϵ</mi> </math></EquationSource> </InlineEquation>-Greedy, and compare their performance in terms of converge quality and speed. We propose an enhanced LRI denoted as LRI <InlineEquation ID="IEq4"> <EquationSource Format="TEX">${\text { sub-strategy }}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mspace width="0.333333em" /> <mtext>sub-strategy</mtext> <mspace width="0.333333em" /> </mrow> </math></EquationSource> </InlineEquation> based on limiting the action set based on previous decisions. We validate the effectiveness of our proposed solution in comparison with the other RL methods through numerical experiments using real-world data.</p>

Performance Investigation of Stateless RL Methods for Optimal Bus Scheduling

期刊论文
发表日期：

2025年7月16日

查看全文

Perla Hajjar,
Leïla Kloul,
Dominique Barth

摘要

This research aims to find a schedule that adapts to the variations in the transport system that influence the traveled time by the buses. For this purpose, the stop-skipping control strategy is adopted with the goal to serve all passengers waiting at stations with a minimum total delay in a well known static environment. We introduce a novel measure to calculate the maximum total delay of passengers based on the notion of balancing the load inside the buses, denoted as ${\text { load-delay }}$ $load-delay$ . To solve the stop-skipping decision problem with minimizing the ${\text { load-delay }}$ $load-delay$ , we model it as a distributed game where the stations are the players. We explore three stateless reinforcement learning algorithms: Linear Reward Inaction (LRI), Upper Confidence Bound (UCB), and $\epsilon $ $ϵ$ -Greedy, and compare their performance in terms of converge quality and speed. We propose an enhanced LRI denoted as LRI ${\text { sub-strategy }}$ $sub-strategy$ based on limiting the action set based on previous decisions. We validate the effectiveness of our proposed solution in comparison with the other RL methods through numerical experiments using real-world data.