wohnjayne / eda-suite

Research code for using generative neural networks in Estimation of Distribution Algorithms (EDAs)

README of eda-suite on github

Overview

1. What is EDA suite
2. How can I use it
3. Publications and configurations
4. Sources and further reading

Contents

##1. What is EDA suite## EDA suite is a tool I use for my research at the University of Mainz, Germany. It is meant as a tool to test different probabilistic models in the context of estimation of distribution algorithms (EDA) (see [1]).

It comprises

• a main EDA file (eda.m) with the basic loop
• a set of benchmark problems (fitnessXXX.m)
• a number of probabilistic models, currently:
• a Restricted Boltzmann Machine (RBM) [2]
• the Bayesian Optimization Algorithm (BOA) [3]
• Population-based incremental learning (PBIL) [4]
• Denoising Autoencoders (DAE) [5]
• Deep Boltzmann Machine (DBM) [6]; (includes some code written by R. Salakhutdinov)
• Generative Adversarial Network (GAN) [7]
• Factored Higher-Order RBM [8]
• some basic selection functions

##2. How can I use it?##

• I used GNU/Octave for all experiments, mostly v3.2.4 and 3.8.1 (you might need to install some extra packages, especially for BOA)
• Start Octave, go to the ./src/ directory
• adjust the EDA parameters (including the choice of problem) in the file set_eda_params.m
• adjust the selection parameters in the file set_selection_params.m
• choose a model which you want to use, and change the appropriate .m-file, e.g. set_rbm_params.m
• load all settings into the ocatave-workspace, and start the EDA, as follows:

set_eda_params;
set_selection_params;
set_ae_params;  %in case you want to use the denoising autoencoder
[mean_fitness, best_fitness, found_in_iteration,cputime_until_best]=eda(eda_params,selection_params,model_params)

##3. Publications and configurations The following publications include experiments performed with this code (see individual comments).

The corresponding configuration files can be found in the directory '.configs-papers/'

Title	Comment	Config subfolder	URL
Model Building and Sampling in Estimation of Distribution Algorithms using Denoising Autoencoders	October 2016 submission, updated version of DAE-EDA	dae-eda	under review
Denoising Autoencoders for Fast Combinatorial Black Box Optimization (GECCO'15 + submission on arXiv)	slightly different autoencoder code was used	gecco15-dae-eda-fast/	ACM and arXiv
Deep Boltzmann Machines in Estimation of Distribution Algorithms for Combinatorial Optimization		dbm-eda/	arXiv
Generative Adversarial Networks in Estimation of Distribution Algorithms for Combinatorial Optimizatoin		gan-eda/	arXiv

##4. Sources and further reading [1] P. Larranaga and J.A. Lozano. Estimation of Distribution Algorithms: A New Tool for Evolutionary Computation, Kluwer Academic Pub, 2002

[2] Hinton, G. E. A Practical Guide to Training Restricted Boltzmann Machines, 2006 Techical Report

[3] Pelikan, M., Hierarchical Bayesian Optimization Algorithm, Springer 2005

[4] Baluja, S. Population-Based Incremental Learning: A Method for Integrating Genetic Search Based Function Optimization and Competitive Learning Carnegie Mellon University, 1994

[5] Vincent, P.; Larochelle, H.; Bengio, Y. & Manzagol, P.-A. Extracting and Composing Robust Features with Denoising Autoencoders Proceedings of the 25th international conference on Machine learning, 2008, 1096-1103

[6] Salakhutdinov, R. & Hinton, G. E. A Better Way to Pretrain Deep Boltzmann Machines Advances in Neural Information Processing Systems, 2012, 2447-2455

[7] Goodfellow, Ian and Pouget-Abadie, Jean and Mirza, Mehdi and Xu, Bing and Warde-Farley, David and Ozair, Sherjil and Courville, Aaron and Bengio, Yoshua; Generative Adversarial Nets; In: Advances in Neural Information Processing Systems (NIPS 2014), 2672-2680

[8] Memisevic, Roland, and Geoffrey E. Hinton. "Learning to represent spatial transformations with factored higher-order Boltzmann machines." Neural Computation 22.6 (2010): 1473-1492.