Increasing the Efficiency of Control of Self-Driving Vehicles Based on Refinement of the Reward Function in Reinforcement Learning

The article considers the urgent problem of efficient interaction of unmanned vehicles in complex road conditions. Despite significant achievements in the field of autonomous driving, there are still unresolved issues related to safety and control efficiency during such driving. The main focus is on reinforcement learning methods, which represent a promising approach to solving this problem. A new approach to the formation of a reward function is proposed, taking into account many factors characteristic of a real road situation. The experiments demonstrate that refining the reward function leads to a stable improvement in the learning performance of agents, which is confirmed by the analysis of the dependencies of metrics on the duration of training. Several quantitative criteria are highlighted, such as the agent's ability to drive safely and minimize potential emergency situations. These criteria are formalized as a partial order, which allows for a more accurate assessment of agent behavior on the road. The results show that agents trained using a refined reward function demonstrate more aggressive and efficient behavior, which can lead to increased traffic efficiency. The conclusion highlights the need for further research in the implementation of reinforcement learning methods to ensure reliable interaction of self-driving cars in real-world conditions. The development of technologies and control methods in this area will contribute to the creation of a safer and more efficient transport system, which is an important task for the future of autonomous driving.