Autonomous traffic signal control using deep reinforcement learning

Abstract

Traffic signals provide one of the primary means to administer conflicting road traffic flows. The efficiency of road transportation systems significantly depends on signal operation. The state of-the-art signal control strategies are unable to efficiently and autonomously adapt to changing traffic flow patterns. In this thesis, we present an autonomous traffic signal control system, in which each intersection independently computes effective signal regimes to optimize traffic flows through that intersection at all instants-based solely on live camera footage. Our signal control system is trained via Deep Reinforcement Learning (DRL). In recent years, DRL has emerged as a powerful paradigm for control optimization problems by autonomously discovering effective control policies. Our signal control agent perceives the traffic situation around an intersection through visual sensory data and continuously modifies the traffic signal regimes in real time, asper the changing traffic observations. The contributions of this thesis are summarised as; (1) A truly adaptive signal control agent, that effectively tailors its signal control decisions to changing traffic patterns and significantly outperforms the conventional signal control methods (both fixed and adaptive) in single and multi-intersection scenarios. (2) This thesis, for the first time, by using transfer learning, empirically demonstrates vision-based signal control agent’s high generalizability and accelerated learning skills on newly-encountered traffic conditions (such as prioritizing the navigation of emergency vehicles, handling adverse weather and lighting conditions). (3)Additionally, this thesis presents the first application of attention-visualization to illustrate the interpretation of DRL agents’ signal control decisions, while highlighting the benefits of using visual traffic data from CCTV cameras for signal control over the conventional traffic data collection methods such as induction loops.