An ant colony optimization algorithm for continuous optimization: application to feed-forward neural network training

Abstract

Ant colony optimization (ACO) is an optimization technique that was inspired by the foraging behaviour of real ant colonies. Originally, the method was introduced for the application to discrete optimization problems. Recently we proposed a first ACO variant for continuous optimization. In this work we choose the training of feed-forward neural networks for pattern classification as a test case for this algorithm. In addition, we propose hybrid algorithm variants that incorporate short runs of classical gradient techniques such as backpropagation. For evaluating our algorithms we apply them to classification problems from the medical field, and compare the results to some basic algorithms from the literature. The results show, first, that the best of our algorithms are comparable to gradient-based algorithms for neural network training, and second, that our algorithms compare favorably with a basic genetic algorithm.

Training neural networks with ant colony optimization algorithms for pattern classification

“…4.3) is used to compare the proposed ACO training with ACO R (Socha and Blum, 2007), an existing ACO training from the literature that does not follow the original ACO framework. The aim of these two experimental studies is to investigate the effect of pheromone trails on the training of feed-forward neural networks.…”

“…A comprehensive survey regarding evolutionary neural networks is available in (Yao, 1999). Recently, swarm intelligence techniques, including PSO (Mendes et al, 2002;Carvalho and Ludermir, 2006), ABC (Karaboga et al, 2007;Karaboga and Ozturk, 2009) and ACO (Socha and Blum, 2007), were also used to train neural networks. A review for other metaheuristics used for training neural networks is available in (Alba and Marti, 2006).…”

“…Some researchers have demonstrated that standalone evolutionary training is faster than gradient descent training (Montana and Davis, 1989), whereas other researchers have demonstrated that there is no any significant difference between the two types of training and the difference really depends on the problem (Socha and Blum, 2007). However, hybrid training with a GA (Alba and Chicano, 2004) or ACO (Liu et al, 2006;Socha and Blum, 2007;Mavrovouniotis and Yang, 2013) usually performs better than standalone metaheuristics or gradient descent algorithms. This is due to the fact that metaheuristics are global optimization algorithms and they are less sensitive on the initial condition of training whereas local optimization algorithms find the local optimum in the neighbourhood of the initial weights given.…”

“…Recently, ACO R was proposed (Socha, 2004) and applied to train feed-forward neural networks since the training process can be considered as a continuous optimization process (Blum and Socha, 2005;Socha and Blum, 2007). ACO R utilizes the continuous probability density function.…”

“…Swarm intelligence algorithms are inspired from the natural behaviour of social insects, such as the ant foraging, bird flocking, fish schooling, and so on. The applications of swarm intelligence for pattern classification include particle swarm optimization (PSO) (Carvalho and Ludermir, 2006;Mendes et al, 2002), artificial bee colony (ABC) (Karaboga and Ozturk, 2009) and ant colony optimization (ACO) (Blum and Socha, 2005;Socha and Blum, 2007). Among these developments, ACO has attracted less attention than PSO and ABC since it was originally introduced to solve discrete optimization problems (Dorigo and Gambardella, 1997;Dorigo and Stützle, 2004) rather than training neural networks, which is a numerical (continuous) optimization problem (Socha, 2004;Socha and Blum, 2007).…”