Benchmarking of neuroevolutionary approach for controlling task of trolley balance

DOI

10.2991/itsmssm-17.2017.7How to use a DOI?

Keywords

Neuroevolution, reinforcement machine learning, optimization, evolutionary computation, fitness function

Abstract

The article analyzes the neuroevolution of the problematic issues - a promising approach for solving complex problems of machine learning neural networks, adaptive management, and multi-agent systems, evolutionary robotics, search game strategies, computer art. The authors propose a neuroevolutionary algorithm that allows to "grow" a neural network for solving the problems of machine learning with reinforcement. The crossover operator is not used in the algorithm. The evolution of the network is performed due to slight mutational changes in a limited area. The advantages of the algorithm include its independence from the type of neuron activation functions, the absence of a training sample, and the ability to automatically find the appropriate neural network architecture. The authors demonstrate the results of benchmarking on the benchmark task, namely, the task of balancing a trolley with two flagpoles of different lengths. The simulation results support the hypothesis of the advantages of generation of neurostructures by small mutational changes in a limited area.

Copyright

© 2017, the Authors. Published by Atlantis Press.

Open Access

This is an open access article distributed under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).

Download article (PDF)

TY  - CONF
AU  - Sergey Rodzin
AU  - Ada Rodzina
PY  - 2017/12
DA  - 2017/12
TI  - Benchmarking of neuroevolutionary approach for controlling task of trolley balance
BT  - Proceedings of the IV International research conference "Information technologies in Science, Management, Social sphere and Medicine" (ITSMSSM 2017)
PB  - Atlantis Press
SP  - 26
EP  - 29
SN  - 2352-538X
UR  - https://doi.org/10.2991/itsmssm-17.2017.7
DO  - 10.2991/itsmssm-17.2017.7
ID  - Rodzin2017/12
ER  -