A Gradient-Guided ACO Algorithm for Neural Network Learning

The ACO-R algorithm is an Ant Colony Optimization (ACO) algorithm for real-valued optimization, and has been applied to neural network learning. Unlike many algorithms for neural network learning, ACO-R does not use gradient information at all in its operation. Also, unlike many discrete ACO algorithms, ACO-R does not allow for the incorporation of domain-specific heuristics. In this work, we present a gradient-guided variation of ACO-R that incorporates gradient information while retaining the core aspects of the ACO-R algorithm. Experimental results using 10-fold cross-validation with 20 UCI datasets indicate that our variation produces lower test set error than standard ACO-R, after a markedly smaller number of training generations.