Controle adaptativo de tráfego: um estudo comparativo entre aprendizado por reforço e semáforos de tempo fixo

Abstract

This work compares intelligent traffic lights, controlled by reinforcement learning, with fixed-time traffic lights according to the Brazilian Traffic Signaling Manual, aiming to evaluate their effectiveness in improving traffic flow and reducing congestion in Brazil. Q-learning and Deep Q-Network (DQN) algorithms were implemented in a simple intersection and a network with three aligned traffic lights. Simulations in the SUMO environment compared the performance of intelligent traffic lights and fixed-time traffic lights in terms of waiting time, number of stops, and average vehicle speed. The results indicated that, in the simple intersection, the DQN system presented a 25% reduction in total stops (7.01 stops) compared to the fixed-time system (9.33 stops). Q-learning presented 8.19 stops. In terms of total waiting time, DQN achieved an average of 39.24 seconds, representing a 58.8% reduction compared to the fixed-time system (95.29 seconds). Q-learning had an average of 66.61 seconds. In the network with three aligned traffic lights, DQN obtained the lowest average total stops (9.74 stops), a reduction of 20.1% compared to the fixed-time system (12.20 stops). Q-learning had an average of 10.96 stops. The total waiting time was lower in DQN (76.94 seconds) compared to the fixed-time system (94.94 seconds) and Q-learning (84.97 seconds). Thus, the implementation of intelligent traffic lights with reinforcement learning can improve traffic management in urban environments.