Designing Convolutional Neural Networks Using a Genetic Approach for Ball Detection

Abstract

At RoboCup 2017, the HULKs reached the Standard Platform League’s quarter finals and won the mixed team competition together with our fellow team B-Human. This paper describes the design of a convolutional neural network used for the detection of the black and white ball - one of the key contributions that led to the team’s success. We present a genetic design approach that optimizes network hyperparameters for a cost effective inference on the NAO, with limited amount of training data. Experimental results demonstrate that the genetic algorithm is able to optimize the hyperparameters of convolutional neural networks. We show that the resulting network is able to run in real-time on the robot with a very precise classification in generalization test.