Skip to content
This project is a part of "Neural Information Processing Project" taught at TU Berlin.
Jupyter Notebook Python
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
Data
notebooks
report
results
scripts
src
trained_models
training-log
.gitignore
DEVELOPMENT.md
README.md
config.yaml
requirements.txt
start-notebook.sh
xfoil-command

README.md

Black Box Optimization using Recurrent Neural Networks

Abstract

We extend on the work of Chen et al., who introduce a new, learning-based, approach to global Black-Box optimization. They train a LSTM model, on functions sampled from a Gaussian process, to output an optimal sequence of sample points, i.e. a sequence of points that minimize those training functions. We verify the claims made by the authors, confirming that such a trained model is able to generalize to a wide variety of synthetic Black-Box functions, as well as to the real world problems of airfoil optimization and hyperparameter tuning of a SVM classifier. We show that the method performs comparably to state-of-the-art Black-Box optimization algorithms on these benchmarks, while outperforming them by a wide margin w.r.t. computation time. We thoroughly investigate the effects of different loss functions and training distributions on the methods performance and examine ways to improve it in the presence of prior knowledge.

Full report and poster can be found at ./report.

Authors

  • Pattarawat Chormai
  • Felix Sattler
  • Raphael Holca-Lammare(Supervisor)

Development Setup

Please check DEVELOPMENT.md.

You can’t perform that action at this time.