MABWiser is a research library written in Python for rapid prototyping of multi-armed bandit algorithms. It supports context-free, parametric and non-parametric contextual bandit models and provides built-in parallelization for both training and testing components. The library also provides a simulation utility for comparing different policies and performing hyper-parameter tuning. MABWiser follows a scikit-learn style public interface, adheres to PEP-8 standards, and is tested heavily.

MABWiser is developed by the Artificial Intelligence Center of Excellence at Fidelity Investments.

# An example that shows how to use the UCB1 learning policy
# to make decisions between two arms based on their expected rewards.

# Import MABWiser Library
from mabwiser.mab import MAB, LearningPolicy, NeighborhoodPolicy

# Data
arms = ['Arm1', 'Arm2']
decisions = ['Arm1', 'Arm1', 'Arm2', 'Arm1']
rewards = [20, 17, 25, 9]

# Model 
mab = MAB(arms, LearningPolicy.UCB1(alpha=1.25))

# Train
mab.fit(decisions, rewards)

# Test
mab.predict()

Available Learning Policies:

• Epsilon Greedy
• LinUCB
• Softmax
• Thompson Sampling (TS)
• Upper Confidence Bound (UCB1)

Available Neighborhood Policies:

• Clusters
• K-Nearest
• Radius

There are two alternatives to install the library:

1. Install from the provided wheel package
2. Build from the source code

The library requires Python 3.6+. The requirements.txt lists the necessary packages. The following packages are used currently:

joblib
numpy
pandas
scikit-learn
scipy
seaborn>=0.9.0

You can install the library from the provided wheel package using the following commands:

git clone https://github.com/fmr-llc/mabwiser.git 
cd mabwiser
pip install dist/mabwiser-X.X.X-py3-none-any.whl

Note: Don't forget to replace X.X.X with the current version number.

Alternatively, you can build a wheel package on your platform from scratch using the source code:

git clone https://github.com/fmr-llc/mabwiser.git
cd mabwiser
pip install setuptools wheel # if wheel is not installed
python setup.py bdist_wheel 
pip install dist/mabwiser-X.X.X-py3-none-any.whl

To confirm that cloning the repo was successful, run the tests and all should pass.

git clone https://github.com/fmr-llc/mabwiser.git
cd mabwiser
python -m unittest discover -v tests

To confirm that installation was successful, import the library in Python shell or notebook.

from mabwiser.mab import MAB, LearningPolicy, NeighborhoodPolicy

To upgrade to the latest version of the library, run git pull origin master in the repo folder, and then run pip install --upgrade --no-cache-dir dist/mabwiser-X.X.X-py3-none-any.whl.

See the /examples folder for self-contained usage examples:

• Content-Free Bandits
• Contextual Parametric Bandits
• Contextual Non-Parametric Bandits
• Customizing Bandits
• Parallelization
• Hyper-Parameter Tuning

Conceptually, given a set of historical decisions and their corresponding rewards, the high-level idea behind MABWiser is to train a model using the fit() method, and then, predict the next best decisions using the predict() method.

Apart from the actual predictions, it is also possible to retrieve the expected reward of each arm using the predict_expectations() method. As time progresses and new decision and reward history becomes available, online training can be performed using the partial_fit() method. New arms can be introduced to the system dynamically using the add_arm() method.

In terms of input data types, decisions and rewards data support lists, 1D numpy arrays, and pandas series while contexts data supports 2D lists, 2D numpy arrays, pandas series and data frames.

There are many real-world situations in which we have to decide between multiple options yet we are only able to learn the best course of action by testing each option sequentially.

Multi-Armed Bandit (MAB) algorithms are suitable for such sequential, online decision making problems under uncertainty. As such, they play an important role in many machine learning applications in internet advertising, recommendation engines, and clinical trials among many others. In this setting, for each and every renewed decision we face an underlying question: Do we stick to what we know and receive an expected result ("exploit") or choose an option we do not know much about and potentially learn something new ("explore")?

Problem Definition: In a multi-armed bandits problem, the model of outcomes is unknown, and the outcomes can be deterministic or stochastic. The agent needs to make a sequence of decisions in time 1, 2, ..., T. At each time t the agent is given a set of K arms, and it has to decide which arm to pull. After pulling an arm, it receives a reward of that arm, and the rewards of other arms are unknown. In a stochastic setting the reward of an arm is sampled from some unknown distribution. Situations exist in which we also observe side information at each time t. This side information is referred to as context. The arm that has the highest expected reward may be different given different contexts. This variant is called contextual multi-armed bandits. Overall, the objective is to minimize regret, or equivalently, maximize the cumulative expected reward in the long run.

MABWiser is licensed under the Apache License 2.0.