daisypy-optim 0.3.0

Optimization framework for Daisy

daisypy-optim

Framework for optimization of Daisy parameters.

Features

• Optimization using a greedy sequential optimizer, CMA-ES, or Bayesian optimizers
• Optimization across multiple scenarios
• Single or multi-objective optimization
• Optimization of parameters in both .dai and .py files

See doc/examples for examples of how to run optimizations.

See https://github.com/orting/interactive-decision-maps for visualization of multi-objective optimization results.

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

To use daisypy-optim as a regular python module. Simply install from pypi with

pip install daisypy-optim

To enable CMA-ES

pip install daisypy-optim[cma]

To enable Ax (this pulls a lot of dependencies)

pip install daisypy-optim[ax]

All of the above

pip install daisypy-optim[cma,ax]

And look at the doc/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.
• ax : Ax. Bayesian optimization. Supports multi-objective optimization.

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). CMA-ES requires continuous parameters. Ax should wotk with both.