Evolving neural networks

Horn, Aidan

Student Work

Evolving neural networks

Public Deposited

NeuroEvolution of Augmenting Topologies (NEAT) is a technique that employs genetic evolution to optimize neural networks to solve a particular machine learning task. The team sought to build upon the capabilities of NEAT to evolve ensembles of neural networks that solve both classification and reinforcement learning tasks. To do so, the team developed a novel fitness function for evaluating a neural network by the performance of ensembles containing it, termed Constituent Ensemble Evaluation. To test this approach, the team developed an ensembling algorithm that maximizes genetic diversity among ensemble members and compared this method to other established algorithms. As ensembles benefit from diversity, the team also attempted to determine an optimal configuration of NEAT parameters to yield a more diverse population of neural networks. It was hypothesized that neural networks evolved using constituent ensemble evaluation and ensembled using a diversity heuristic would outperform ensembles generated using individual fitness. For simple reinforcement learning tasks, constituent ensemble evaluation did not improve the capability for NEAT to evolve useful ensembles, regardless of the ensembling method used. For classification tasks, the team claims that evolving individuals based on their constituent ensemble performance is not sufficient alone; however, when constituent ensemble evaluation is incrementally introduced into the fitness function, ensembles yield better overall performance. This approach yielded an average test accuracy of 0.601 ± 0.1 on the UCI Heart Disease data set, which outperforms an alternative method for evolving ensembles, Orthogonal Evolution of Teams.

This report represents the work of one or more WPI undergraduate students submitted to the faculty as evidence of completion of a degree requirement. WPI routinely publishes these reports on its website without editorial or peer review.

Creator