Ajuste automático de parâmetros de meta-heurísticas com aprendizado de máquina

Abstract

The Traveling Salesman Problem (TSP) stands as a cornerstone challenge in combinatorial optimization, with widespread applications in logistics, route planning, and network management. The performance of metaheuristic algorithms, such as Ant Colony Optimization, hinges critically on precise parameter tuning — a task that is both intricate and computationally demanding due to the inherent variability across problem instances. This work presents OptAI, a Python-based library designed for automated parameter optimization in the TSP, serving as a case study. OptAI seamlessly integrates advanced Machine Learning (ML) techniques — including XGBoost, Random Forest, and SVR — with robust optimization frameworks like IRACE and Optuna. By analyzing historical logs from IRACE, the library trains regression models to predict instance-specific parameter configurations, leveraging descriptive features such as the number of cities, density, and distance statistics. In contrast to IRACE, which provides a single parameter set for all instances, OptAI delivers tailored configurations by harnessing the predictive capabilities of ML models. Furthermore, the library enhances scalability by circumventing exhaustive parameter space exploration and supports modular integration with modern frameworks for continuous updates. Preliminary results demonstrate its efficacy, although its reliance on IRACE logs currently limits its autonomy. OptAI marks a significant advancement by bridging artificial intelligence with classical methods, offering a hybrid approach that combines Machine Learning and metaheuristics to tackle complex combinatorial optimization problems effectively.