Curators

A package for action selection and optimization.

Action types currently supported are context-naive, contextual, and sequential.

• Context-naive actions, which we'll call simply "naive" assumes the expected return is simply a function of the action itself. This is commonly referred to as a multi-armed bandit.
• Contextual actions enable use of observable variables to further improve the accuracy of the expected return of actions. For now, we enabled integration of clustering algorithms with a traditional bandit. Longer term, this will grow into more broad support for contextual bandits.
• Sequential algorithms extend the previous optimizations of single actions into the optimization of multiple ordered actions.

To facilitate maintaining the mathematical purity of classic models while maintaining practical interfaces for real world actions, objects are categorized into three object types: models, agents, and applications.

• Model - distinct and mathematically pure units of learning.
• Agent - integration of identically shaped models into a single interface.
• Application - tooling around connecting an agent to some real-world medium.

Installation

Pip

pip install curators

Package Build and Install

We use Python's setuptools to handle installation. You can build the wheel and install it elsewhere manually using:

# from within this repo's root:
python setup.py bdist_wheel

# from within your other project that you're installing this wheel into:
pip install curators-0.0.1-py3-none-any.whl

Or you can install it directly from Github usisng the library name git+ssh://git@github.com/BellwethrInc/curators.git as you would install any other library through either a pip install [x] or including it directly in your requirements.txt file.

Or if you prefer using https rather than ssh to connect to github, git+https://github.com/BellwethrInc/curators.git works as well.

After we get proper releases set up, you'll be able to do git+ssh://git@github.com/BellwethrInc/curators.git@v0.0.0#egg=curators to specify a release as well.

Coming soon:

• proper versioning

As a Developer

Within this repo's root, create a virtualenv (or your conda equivalent) and install the dependencies.

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Testing

Within a developer's environment, run all tests simply with

    python -m unittest

Coverage

To see testing coverage, run

    coverage run -m unittest discover  # to generate coverage data
    coverage report  # to render the coverage data to console
    coverage-badge -o docs/images/coverage.svg  # update the coverage badge

Docs

To build docs

• within your venv
• pip install -r requirements.txt
• cd docs
• make html

Or if you're within pycharm:

• be sure your current environment also has the docs requirements installed
  • pip install -r docs/requirements.txt
• Create and configure a new "Spinx Task" configuration:
  • Be sure html is the command selected
  • select docs/source as the source directory
  • select docs/build directory.
    • Make this directory yourself or let sphinx make it for you.
    • this path is within .gitignore so don't worry about stepping on toes within your build dir.
• rerun docs build as needed
• docs/build/index.html is the file you want to open in your browswer

Notebooks

When running the notebooks, you need to either be working within a virtualenv with curators installed or have the root of this repo within your environment's PATH or PYTHONPATH. For example,

PYTHONPATH="${PYTHONPATH}:path/to/this/repos/root" jupyter notebook

will leave the rest of your environments unaffected while giving python access to this repo directly.