Temporal Action Analysis in Metaheuristics: A Machine Learning Approach

Abstract

This study explores the use of Autoregressive Integrated Moving Average (ARIMA) and Long Short-Term Memory (LSTM) machine learning models in metaheuristic algorithms, with a focus on a modified General Variable Neighborhood Search (GVNS) for the Capacitated Vehicle Routing Problem (CVRP). We analyze the historical chain of actions in GVNS to demonstrate the predictive potential of these models for guiding future heuristic applications or parameter settings in metaheuristics such as Genetic Algorithms (GA) or Simulated Annealing (SA). This “optimizing the optimizer” approach reveals that, the history of actions in metaheuristics provides valuable insights for predicting and enhancing heuristic selections. Our preliminary findings suggest that machine learning models, using historical data, offer a pathway to more intelligent and data-driven optimization strategies in complex scenarios, marking a significant advancement in the field of combinatorial optimization.