Designing Convolutional Neural Networks Using a Genetic Approach for Ball Detection

First published in
Lecture notes in computer science  
Number in series
11374 LNCS
Start Page
150
End Page
161
Citation
Lecture Notes in Computer Science (11374 LNCS): 150-161 (2019)
Contribution to Conference
Robot World Cup, RoboCup 2018  
Publisher DOI
10.1007/978-3-030-27544-0_12
Scopus ID
2-s2.0-85070719936
Publisher
Springer
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.
DDC Class
004: Informatik