sweetpea 0.2.5

A language for synthesizing randomized experimental designs

SweetPea

Build Status

SweetPea is a domain-specific language for specifying factorial experimental designs and synthesizing trial sequences from those design specifications. An explanation of factorial experimental designs and how to build and manipulate them in SweetPea can be found in the SweetPea Guide.

See also the paper describing SweetPea, but beware that the API has changed. The main changes are replacing most functions with classes, changing some function names, and simplifying the organization to just a sweetpea module that exports everything.

SweetPea includes a synthesizer to generate sequences of trials that satisfy the design's constraints. The goal is to generate sequences that are unbiased: every possible sequence of trials that satifies the design constraints is equally likely to be generated, which avoids correlations that are not part of the experiment's definition. With current technologies, SweetPea achieves that goal for designs with either simple constraints or a small number of combinations. SweetPea can also generate samples that seem uniform in practice for medium-sized designs, although without a formal guarantee. Generating unbiased samples for large designs remains an area of active research and development.

For designs that do not involve constraints that span trials within a sequence, SweetPea can directly sample with combinatoric techniques. Realistic designs often involve transition constraints or other cross-trial constraints, however. For those cases, SweetPea's primary sampling strategy compiles an experiment design into a boolean formula; compilation ensures a 1-to-1 correspondence between distinct satisfying assignments to the boolean formula and distinct trial sequences, so that uniformly sampling solutions to the boolean formula imples a unform sample of trial sequences. SweetPea uses CMSGen and UniGen to sample solutions to the boolean formula. CMSGen generates samples that appear to be well distributed in practice, but CMSGen lacks a formal guarantee of uniformity. UniGen provides statistical guarantees that the solutions it finds are approximately uniformly probable, but its approach is tractable only for the smallest designs that are expressed with SweetPea.

Dependencies

SweetPea requires Python 3.7.9 or later.

Installation

There are two ways to install SweetPea: from the Python Package Index (PyPI), or from source.

Installing from PyPI

SweetPea can be installed from PyPI via pip:

$ pip install sweetpea

Installing from Source

Clone this repository, install SweetPea's dependencies, and install SweetPea itself:

$ git clone https://github.com/sweetpea-org/sweetpea-py.git
$ cd sweetpea-py
$ pip install -r requirements.txt
$ pip install .

---

:exclamation: Important!

The pip install . command installs SweetPea locally, but it will not automatically check for updates. If you intend to manually update your local copy of SweetPea, you should instead do pip install -e . to tell pip to use the source dynamically.

---

Examples

There are example programs in the example_programs directory, and there is a detailed explanation of how to use SweetPea in the SweetPea Guide.

API Documentation

The SweetPea API is documented in the API section of the SweetPea Guide.

Contributing

Information on how to contribute to SweetPea's development can be found in the Contributing section of the SweetPea Guide.