Advanced

Optimering av hyperparametrar till artificiella neurala nätverk med genetiska algoritmer

Stensson, Simon (2016) FMS820 20161
Mathematical Statistics
Abstract
This master thesis explores the feasibility of using genetic algorithms in order to automate the process of optimizing hyperparameters for artificial neural networks (ANN). Today there is no standard way to optimize hyperparameters for ANN; often they are set manually by trial and error. In order to explore the feasibility of using genetic algorithms to optimize hyperparameters for ANN, two algorithms are implemented in Python. The first is a genetic algorithm and the second is an algorithm that trains a neural network and enables predictions. The two algorithms interact in a feedback loop where the genetic algorithm adjusts hyperparameters for the neural network, and the neural network performs predictions on data which is used as... (More)
This master thesis explores the feasibility of using genetic algorithms in order to automate the process of optimizing hyperparameters for artificial neural networks (ANN). Today there is no standard way to optimize hyperparameters for ANN; often they are set manually by trial and error. In order to explore the feasibility of using genetic algorithms to optimize hyperparameters for ANN, two algorithms are implemented in Python. The first is a genetic algorithm and the second is an algorithm that trains a neural network and enables predictions. The two algorithms interact in a feedback loop where the genetic algorithm adjusts hyperparameters for the neural network, and the neural network performs predictions on data which is used as feedback to the genetic algorithm. In order to evaluate the method, the implemented models are evaluated on three classification problems. The results are compared to predictions made from neural networks where the hyperparameters are manually set. The method using genetic algorithms to optimize hyperparameters performes slightly better on all three problems, but without a significant improvement in prediction accuracy. The implemented models offer an automated way to optimize hyperparameters for ANN and test results indicates that prediction accuracy is maintained. (Less)
Please use this url to cite or link to this publication:
author
Stensson, Simon
supervisor
organization
course
FMS820 20161
year
type
H2 - Master's Degree (Two Years)
subject
language
Swedish
id
8884763
date added to LUP
2016-06-23 10:55:49
date last changed
2016-06-23 10:55:49
@misc{8884763,
  abstract     = {This master thesis explores the feasibility of using genetic algorithms in order to automate the process of optimizing hyperparameters for artificial neural networks (ANN). Today there is no standard way to optimize hyperparameters for ANN; often they are set manually by trial and error. In order to explore the feasibility of using genetic algorithms to optimize hyperparameters for ANN, two algorithms are implemented in Python. The first is a genetic algorithm and the second is an algorithm that trains a neural network and enables predictions. The two algorithms interact in a feedback loop where the genetic algorithm adjusts hyperparameters for the neural network, and the neural network performs predictions on data which is used as feedback to the genetic algorithm. In order to evaluate the method, the implemented models are evaluated on three classification problems. The results are compared to predictions made from neural networks where the hyperparameters are manually set. The method using genetic algorithms to optimize hyperparameters performes slightly better on all three problems, but without a significant improvement in prediction accuracy. The implemented models offer an automated way to optimize hyperparameters for ANN and test results indicates that prediction accuracy is maintained.},
  author       = {Stensson, Simon},
  language     = {swe},
  note         = {Student Paper},
  title        = {Optimering av hyperparametrar till artificiella neurala nätverk med genetiska algoritmer},
  year         = {2016},
}