Prestandajämförelse av pathfinding algoritmer i gatukartor : En studie av LLMA* och Bidirectional Dijkstra

Sammanfattning

This research paper explores alternative applications for Large Language Models (LLMs) beyond traditional text and code generation. An artifact was developed to perform pathfinding within a localized street map, capable of utilizing various algorithms. The study focuses on comparing Bidirectional Dijkstra with LLMA*, a novel approach proposed by Meng et al. that leverages an LLM to identify checkpoints, which are subsequently connected using an A* algorithm. This methodology aims to exploit the efficiency of A* in restricted search spaces while extending its reach over larger geographical areas. The experiment involved 100 iterations for each of 10 randomly selected node pairs. Performance was evaluated based on execution time, node exploration count, CPU and RAM utilization, and path length. The results indicate that LLMA* is significantly slower than Bidirectional Dijkstra and produces suboptimal paths if any.