daisypy-optim 0.1.0

Optimization framework for Daisy

daisypy-optim

Framework for optimization of Daisy parameters.

Overview

The framework can be used in two ways

1. As a self-contained environment
2. As a regular python module

Self contained optimization environment

The idea of self-contained environment is to create a repository that contains everything that is needed for the optimization (except the Daisy binary). This allow easy sharing, version control and archiving. This type of optimization is defined by

• a single dai file template
• zero or more python file templates
• a python script that runs the simulation

A tool is included that generates the python script based on user input. Once generated, it can be adapted to specific needs. To start using daisypy-optim this way look at

• Windows
• Linux

Regular python modules

It is also possible to use daisypy-optim as a regular python module. Simply install with

pip install daisypy-optim@git+https://github.com/daisy-model/daisypy-optim

To enable CMA-ES

pip install cma

To enable scikit-optimize

pip install scikit-optimize joblib

All of the above

pip install daisypy-optim@git+https://github.com/daisy-model/daisypy-optim cma scikit-optimize joblib

And look at the examples and tests to see how to setup an optimization

Chossing an optimization method

There are several optimization methods available

• sequential : A greedy approach that fixes one parameter at a time. Works well when optimizing few paramaters.
• cma : CMA-ES. Works well for larger number of parameters and when the outcome depends strongly on parameter combinations. Does not work for single parameters.
• skopt : scikit-optimize. Bayesian optimization. Very slow. Will be replaced with a BoTorch/Ax method.

Continuous/categorical parameters

Both continuous and categorical parameters are supported. Currently, the sequential method treats everything as categorial parameters (by sampling a fixed number of values from the continuous parameters). The other methods requires continuous parameters.