Autonomous Driving Decision Model Based on Reinforcement Learning Algorithm in the Environment of Intelligent Connected Vehicles

Existing intelligent perception and control technologies still have problems such as poor generalization, rough control, and unstable response in terms of decisionmaking accuracy, detailed behavior control, and multi-objective trade-offs. To this end, this paper introduces the deep reinforcement learning method TD3 (Twin Delayed DDPG) in artificial intelligence, integrates the IoT perception environment, and constructs an autonomous driving decision model based on continuous action space. By modeling the state space in detail and combining multi-objective reward functions to perform weighted optimization of safety, traffic efficiency, and comfort, a higher-granularity behavioral decision control is achieved. Experimental results show that compared with DDPG, SAC, and traditional rule strategies, the proposed model reduces the collision rate by more than 70 % under typical working conditions, and significantly improves the detailed control capability and overall performance of the autonomous driving system in complex traffic scenarios.