Left Turns at Unsignalized Intersections: Decision Making for Autonomous Vehicles Using Deep Reinforcement Learning with Game-Theoretic Reasoning

Abstract

Left-turn maneuvers at unsignalized intersections pose significant challenges for autonomous vehicles (AVs) as a result of the interactive and uncertain nature of gap-acceptance decisions. This study proposes a deep reinforcement learning-based left-turn with game-theoretic reasoning (DLT-GT) framework that embeds strategic interaction modeling into learning-based decision-making. The framework is formulated as a general integration architecture that can be coupled with different deep reinforcement learning (DRL) agents, with proximal policy optimization serving as a stable implementation backbone in this study. The framework integrates discrete strategic decisions and continuous control through a risk-weighted payoff structure that explicitly captures interactions with surrounding vehicles and pedestrians. Microscopic traffic simulation experiments evaluate performance under varying traffic volumes, aggressiveness levels, and safety sensitivity thresholds. Results show that incorporating game-theoretic reasoning substantially improves decision stability, efficiency, and safety compared with a baseline DRL agent, achieving higher rewards, shorter maneuver times, and zero collisions across all scenarios. These findings demonstrate that augmenting DRL with explicit interaction reasoning provides a generalizable pathway toward safer and more adaptive AV decision-making in complex multiagent environments.