Vehicle Decision-Making System Based on Knowledge Graph and Large Model

The decision-making of vehicles in complex environments is a critical aspect of autonomous driving technology. To address issues such as poor generalization and insufficient adaptability to dynamic environments in traditional decision-making methods, an intelligent decision-making framework that integrates knowledge graph and large language model (LLM) was constructed. The system first collected dynamic data on vehicles and their environment using the CARLA simulation platform, which was then combined with pre-defined static knowledge graphs to form a continuously updated comprehensive knowledge graph, uniformly stored in the Neo4j database. Subsequently, Cypher rule inference quickly generated prior decision candidates; on this basis, LLM was introduced for knowledge-enhanced retrieval and chain-of-thought reasoning, achieving optimization and explanation of strategies. Experimental results show that in simulated urban environments, the comprehensive decision accuracy of knowledge graph-based decision-making was over 95%, with an average response time of less than 60 ms; in off-road vehicle experiments, compared to baselines using only knowledge graphs or only LLM, the proposed method showed improved accuracy. This validates the complementary advantages of structured knowledge from knowledge graphs and LLM inference capabilities, providing a viable path for highly reliable decision-making in autonomous vehicles.