Comparative Study of LSTM and Transformer-Based Reinforcement Learning for Traffic Signal Control

Addressing urban traffic congestion, this paper presents a reinforcement learning (RL)-based traffic signal control (TSC) framework, comparing the effectiveness of Long Short-Term Memory (LSTM) and Transformer architectures for temporal feature extraction. A four-way intersection model is developed, where a Proximal Policy Optimization (PPO) framework is employed with vehicle occupancy ratio as the state, phase switching as the action, and average waiting time as the reward. The proposed models are evaluated through simulations on the SUMO platform. Experimental results indicate that in simple traffic scenarios, the LSTM-based model achieves a significantly lower average waiting time (6.97s) compared to the Transformer-based approach (10.54s) and traditional fixed-phase control (83.13s). These findings highlight the superiority of sequential models in environments dominated by short-term dependencies. Furthermore, this study provides novel insights into the applicability of attention mechanisms versus recurrent neural networks, offering valuable guidance for intelligent traffic systems.