Home // ICONS 2019, The Fourteenth International Conference on Systems // View article

Monte Carlo Tree Search for Optimizing Hyperparameters of Neural Network Training

Authors:
Karolina Polanska
Wiktoria Dywan
Piotr Labuda
Leszek Koszalka
Iwona Pozniak-Koszalka
Andrzej Kasprzak

Keywords: algorithm; Monte Carlo approach; Tree search; hyperparameter; neural network

Abstract:
In tasks related to machine learning, the right selection of hyper-parameters can significantly impact training time and quality of the obtained results. Often, iterative search algorithms are used. In this paper, we propose an approach, based on our own modification of Monte Carlo Tree Search. The new algorithm is designed to work on discrete hyper-parameter spaces, and uses feedback from training process to learn and adjust its subsequent outputs. In the paper, the properties of the algorithm are studied, in particular for training Multilayer Perceptron. Moreover, three search algorithms are compared: Grid Search, Random Search and the proposed Monte Carlo Tree Search. As it is shown, the Monte Carlo Tree Search can give promising results and can be treated as fair competition to the off-shelf solutions.

Pages: 61 to 64

Copyright: Copyright (c) IARIA, 2019

Publication date: March 24, 2019

Published in: conference

ISSN: 2308-4243

ISBN: 978-1-61208-696-5

Location: Valencia, Spain

Dates: from March 24, 2019 to March 28, 2019