Evolving Indirectly Encoded Convolutional Neural Networks to Play Tetris With Low-Level Features

Title

Evolving Indirectly Encoded Convolutional Neural Networks to Play Tetris With Low-Level Features

Creator

Schrum, Jacob

Date

2018-12-10

Date Available

2018-12-10

Date Issued

2018-07

Identifier

Jacob Schrum. 2018. Evolving indirectly encoded convolutional neural networks to play tetris with low-level features. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO '18), Hernan Aguirre (Ed.). ACM, New York, NY, USA, 205-212. DOI: https://doi.org/10.1145/3205455.3205459

uri

https://collections.southwestern.edu/s/suscholar/item/240

Abstract

Tetris is a challenging puzzle game that has received much attention from the AI community, but much of this work relies on intelligent high-level features. Recently, agents played the game using low-level features (10 X 20 board) as input to fully connected neural networks evolved with the indirect encoding HyperNEAT. However, research in deep learning indicates that convolutional neural networks (CNNs) are superior to fully connected networks in processing visuospatial inputs. Therefore, this paper uses HyperNEAT to evolve CNNs. The results indicate that CNNs are indeed superior to fully connected neural networks in Tetris, and identify several factors that influence the successful evolution of indirectly encoded CNNs.

Language

English

Publisher

Proceedings of the Genetic and Evolutionary Computation Conference

Subject

Computing methodologies

Machine learning

Machine learning approaches

Neural networks

Bio-inspired approaches

Generative and developmental approaches

Type

Article