When Memory Matters: An Evaluation of LSTM-Based Multi-Agent Learning for Multi-Intersection Traffic Signal Control
White Rose Research Online (University of Leeds, The University of Sheffield, University of York) · 2026
Abstract
Temporal memory modules such as LSTMs are widely adopted in multi-agent reinforcement learning (MARL) for adaptive traffic signal control (TSC). Yet it remains unclear when such memory is genuinely beneficial and when its added complexity outweighs its gains. In this paper, we conduct a structured diagnostic study that analyses the role of temporal memory in cooperative TSC, disentangling temporal, spatial, and centralised-training inductive biases across value-based and policy-based paradigms under controlled architecture and demand conditions. Our analysis shows that LSTM-based memory provides selective benefits when traffic demand ex- hibits reproducible temporal structure, especially when paired with spatial message passing through GNNs. However, these gains are not universal and do not consistently increase with network size. In contrast, GNN-based spatial encoding provides the most robust architectural benefit across larger networks and structured demand regimes. Under centralised training with decentralised execution (CTDE), recurrence is strongly regime-dependent: it often slows early learning and increases variance while offering limited or inconsistent improvement in final performance. We provide controlled evidence for these effects in throughput, delay, and learning stability, and derive deployment-oriented design rules indicating when temporal memory can improve traffic performance and when it adds complexity without clear benefit.
Citation
Yingyi Kuang, George Vogiatzis, STEVEN XIAOTIAN DAI and Maria Chli. “When Memory Matters: An Evaluation of LSTM-Based Multi-Agent Learning for Multi-Intersection Traffic Signal Control.” White Rose Research Online (University of Leeds, The University of Sheffield, University of York). 2026.
BibTeX
@article{kuang2026,
title = {When Memory Matters: An Evaluation of LSTM-Based Multi-Agent Learning for Multi-Intersection Traffic Signal Control},
author = {Yingyi Kuang and George Vogiatzis and STEVEN XIAOTIAN DAI and Maria Chli},
journal = {White Rose Research Online (University of Leeds, The University of Sheffield, University of York)},
year = {2026},
}