Game-theoretic Decision-making for Autonomous Driving Vehicles

In the next 50 years, self-driving cars will coexist with human-driven vehicles on roads, resulting in constant interactions between the two types of vehicles. It is challenging to solve decision-making problems for autonomous driving vehicles in dynamic and interactive environments. Predefining rules via hard coding is almost impossible to address these issues. In this research, we aim to utilize a combination of game theory and deep reinforcement learning to address the interaction challenge faced by autonomous vehicles. Our approach involves leveraging learning vehicles to explore complex interactions that human drivers experience in real-life situations using a realistic simulator. As the policy evolves, autonomous driving vehicles can drive more successfully by collecting data through interactions with multiple vehicles from the virtual environments. To enable autonomous driving vehicles to make decisions in traffic scenarios with varying levels of reasoning, we have integrated the learning capability of deep q networks with the Level-k reasoning approach. Therefore, vehicles can learn complex driving behaviors via a self-play training mode. The trained policies were tested in different traffic scenarios including unsignalized intersections and urban lane-change scene with different driving models, and the experimental results show that 270the autonomous driving vehicles can still drive safely in such complex traffic environments, thus verifying the effectiveness of our proposed method.