แนวคิดหลัก
The proposed AEL framework combines the power of evolutionary computation and large language models to automatically design highly competitive guided local search algorithms for solving the traveling salesman problem, outperforming human-designed algorithms.
บทคัดย่อ
The paper presents a novel framework called Algorithm Evolution using Large Language Model (AEL) that combines evolutionary computation and large language models to enable automatic algorithm design. The authors demonstrate the effectiveness of this approach by using AEL to design an efficient guided local search (GLS) algorithm for solving the traveling salesman problem (TSP).
The key highlights are:
The AEL framework maintains a population of algorithms and utilizes evolutionary operations like crossover and mutation, facilitated by large language models, to automatically evolve new algorithms.
The evolved GLS algorithm, named AEL-GLS, is evaluated on 1,000 randomly generated TSP instances of sizes 20, 50, and 100, as well as 29 instances from the TSPLib benchmark.
The results show that AEL-GLS outperforms state-of-the-art human-designed GLS algorithms, achieving a 0% gap on TSP20 and TSP50 instances, and a 0.032% gap on TSP100 instances, all within 1,000 iterations.
Compared to recent deep learning-based solvers, AEL-GLS demonstrates superior generalization, as the algorithm evolved on TSP100 instances can be directly applied to smaller problem sizes without retraining.
The entire algorithm evolution process takes only two days, with minimal human effort and no model training required, highlighting the potential of the AEL framework to usher in a new era of automatic algorithm design.
สถิติ
The average gap (%) to the optimal solution obtained by the Concorde solver on the test instances:
TSP20: 0.0%
TSP50: 0.0%
TSP100: 0.032%
คำพูด
"In just two days, the GLS algorithm evolved by AEL achieved superior performance compared to human-designed GLS algorithms and many other algorithms developed over the past decades."
"The results demonstrate that the algorithm generated through AEL surpasses the manually crafted ones, achieving a 0% gap on TSP20 and TSP50, and a 0.032% gap on TSP100 in 1,000 iterations."