Initializing Ant Colony Algorithms by Learning from the Difficult Problem’s Global Features

Abstract

Deception, which stems from the tackled problem instance and algorithmic structure, has a tremendous negative impact on the algorithmic performance. An improved ACO called GFL-ACO with a global feature learning strategy is proposed to process the algorithmic initialization. The strategy consists of two parts: a greedy random walking of ant colony and a mean value approach. With the former part, some initialized ants are launched to step forward by a greedy rule till finished a tour. A statistical manner of edge-based relative frequency is used to initial pheromone trails and ants’ starting positions. With the latter part, a mean value calculated from edge-based relative frequency is used to generate ant population size. The experiments on the TSPLIB benchmark show that GFL-ACO can achieve a rather better performance on the standard benchmark.