coba 4.5.1

A contextual bandit research package.

Coba

What is it?

Coba is a powerful research framework built specifically for contextual bandit (CB) methods.

How do you research?

Think for a second about the last time you considered evaluating CB methods for a research project

1. Was it easy to incorporate your own data sets during CB analysis?
2. Was it easy to incorporate the wide array of available CB algorithms?
3. Was it easy to create, run and share the results of CB experiments?

Coba was built from the ground up to do these three things and more.

The Coba Way

Coba is...

• ... light-weight (it has no dependencies to get started)
• ... distributed (it was built to work across the web with caching, api-key support, checksums and more)
• ... verbose (it has customizable, hierarchical logging for meaningful, readable feedback on log running jobs)
• ... robust (experiments write every action to file so they can always be resumed if stopped for any reason)
• ... just-in-time (no resources are loaded until needed, and they are released immediately to keep memory small)
• ... a duck? (Coba relies only on duck-typing so no inheritance is needed to extend coba's built-in functionality)

But don't take our word for it. We encourage you to look at the code yourself or read more below.

Workflow

Coba is architected around a simple workflow: Learners -> Environments -> Experiments -> Results -> Analysis.

Environments represent unique CB problems that need to be solved. Learners are the CB algorithms that we can use to learn policies. Experiments are combinations of Environments and Learners that we want to evaluate. And Results are the outcome of an Experiment, containing all the data from the Experiment.

Learners

Learners are algorithms which are able to improve their action selection through interactions with environments.

A number of algorithms are provided natively with Coba for quick comparsions. These include:

• All contextual bandit learners in VowpalWabbit
• UCB1-Tuned Bandit Learner by Auer et al. 2002
• LinUCB by Chu et al. 2011
• Corral by Agarwal et al. 2017

Environments

Environments are the core unit of evaluation in Coba. They are nothing more than a sequence of interactions with contexts, actions and rewards. A number of tools have been built into Coba to make simulation creation easier. All these tools are defined in the coba.environments module. We describe a few of these tools here.

Creating Environments From Classification Data Sets

Classification data sets are the easiest way to create Environments in Coba. Coba natively supports:

• Binay, multiclass and multi-label problems
• Dense and sparse representations
• Openml, Csv, Arff, Libsvm, and the extreme classification (Manik) format
• Local files and files over http (with local caching)

The classification environments built into Coba are OpenmlSimulation, CsvSimulation, ArffSimulation, LibsvmSimulation, and ManikSimulation.

Creating Environments From Generative Functions

Sometimes we have well defined models that an agent has to make decisions within but no data. To support evaluation in these domains one can use LambdaSimulation to define generative functions for that will create an Environment.

Creating Environments From Scratch

If more customization is needed beyond what is offered above then you can easily create your own simulation by implementing Coba's simple SimulatedEnvironment interface.

Experiments

Experiments are combinations of Learners and Environments. The Experiment class orchestrates the complex tasks of evaluating learners on environments in a resource efficient, repeatable, and robust manner. Experiments can be run on any number of processes (we often run on 96 core machines) and writes all results to file as it runs so that it can be restarted if it ever stops.

Results

The end result of an Experiment is a Result object which has been built for easy analysis. The result class provides a number of plots as well as filtering capabilities out of the box. If the built in functionality isn't advanced enough data from experiments can be accesed directly from a three table structure: a Learner table, Environment table, and Interaction table. These tables can be ported to pandas dataframes where all manner of advanced filtering and plotting can be performed to discover the strengths and weaknesses of Learners on various Environments.

Examples

An examples directory is included in the repository with a number of code and experiment demonstrations. These examples show how to create experiments, evaluate learners against them and plot the results.