Skip to main content

A multi-armed bandit library for Python

Project description

slots

A multi-armed bandit library for Python

Slots is intended to be a basic, very easy-to-use multi-armed bandit library for Python.

PyPI PyPI - Python Version Downloads

Code style: black type hints with mypy

Author

Roy Keyes -- roy.coding@gmail

License: MIT

See LICENSE.txt

Introduction

slots is a Python library designed to allow the user to explore and use simple multi-armed bandit (MAB) strategies. The basic concept behind the multi-armed bandit problem is that you are faced with n choices (e.g. slot machines, medicines, or UI/UX designs), each of which results in a "win" with some unknown probability. Multi-armed bandit strategies are designed to let you quickly determine which choice will yield the highest result over time, while reducing the number of tests (or arm pulls) needed to make this determination. Typically, MAB strategies attempt to strike a balance between "exploration", testing different arms in order to find the best, and "exploitation", using the best known choice. There are many variation of this problem, see here for more background.

slots provides a hopefully simple API to allow you to explore, test, and use these strategies. Basic usage looks like this:

Using slots to determine the best of 3 variations on a live website.

import slots

mab = slots.MAB(3, live=True)

Make the first choice randomly, record responses, and input reward 2 was chosen. Run online trial (input most recent result) until test criteria is met.

mab.online_trial(bandit=2,payout=1)

The response of mab.online_trial() is a dict of the form:

{'new_trial': boolean, 'choice': int, 'best': int}

Where:

  • If the criterion is met, new_trial = False.
  • choice is the current choice of arm to try.
  • best is the current best estimate of the highest payout arm.

To test strategies on arms with pre-set probabilities:

# Try 3 bandits with arbitrary win probabilities
b = slots.MAB(3, live=False)
b.run()

To inspect the results and compare the estimated win probabilities versus the true win probabilities:

# Current best guess
b.best()
> 0

# Estimate of the payout probabilities
b.est_probs()
> array([ 0.83888149,  0.78534031,  0.32786885])

# Ground truth payout probabilities (if known)
b.bandits.probs
> [0.8020877268854065, 0.7185844454955193, 0.16348877912363646]

By default, slots uses the epsilon greedy strategy. Besides epsilon greedy, the softmax, upper confidence bound (UCB1), and Bayesian bandit strategies are also implemented.

Regret analysis

A common metric used to evaluate the relative success of a MAB strategy is "regret". This reflects that fraction of payouts (wins) that have been lost by using the sequence of pulls versus the currently best known arm. The current regret value can be calculated by calling the mab.regret() method.

For example, the regret curves for several different MAB strategies can be generated as follows:

import matplotlib.pyplot as plt
import slots

# Test multiple strategies for the same bandit probabilities
probs = [0.4, 0.9, 0.8]

strategies = [{'strategy': 'eps_greedy', 'regret': [],
               'label': '$\epsilon$-greedy ($\epsilon$=0.1)'},
              {'strategy': 'softmax', 'regret': [],
               'label': 'Softmax ($T$=0.1)'},
              {'strategy': 'ucb', 'regret': [],
               'label': 'UCB1'},
              {'strategy': 'bayesian', 'regret': [],
               'label': 'Bayesian bandit'},
              ]

for s in strategies:
 s['mab'] = slots.MAB(probs=probs, live=False)

# Run trials and calculate the regret after each trial
for t in range(10000):
    for s in strategies:
        s['mab']._run(s['strategy'])
        s['regret'].append(s['mab'].regret())

# Pretty plotting
plt.style.use(['seaborn-poster','seaborn-whitegrid'])

plt.figure(figsize=(15,4))

for s in strategies:
    plt.plot(s['regret'], label=s['label'])

plt.legend()
plt.xlabel('Trials')
plt.ylabel('Regret')
plt.title('Multi-armed bandit strategy performance (slots)')
plt.ylim(0,0.2);

Regret plot

API documentation

For documentation on the slots API, see slots-docs.md.

Todo list

  • More MAB strategies
  • Argument to save regret values after each trial in an array.
  • TESTS!

Contributing

I welcome contributions, though the pace of development is highly variable. Please file issues and submit pull requests as makes sense.

The current development environment uses:

  • pytest >= 5.3 (5.3.2)
  • black >= 19.1 (19.10b0)
  • mypy = 0.761

You can pip install these easily by including dev-requirements.txt.

For mypy config, see mypy.ini. For black config, see pyproject.toml.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

slots-0.4.0.tar.gz (174.8 kB view details)

Uploaded Source

Built Distribution

slots-0.4.0-py3-none-any.whl (8.7 kB view details)

Uploaded Python 3

File details

Details for the file slots-0.4.0.tar.gz.

File metadata

  • Download URL: slots-0.4.0.tar.gz
  • Upload date:
  • Size: 174.8 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/3.1.1 pkginfo/1.5.0.1 requests/2.22.0 setuptools/45.0.0 requests-toolbelt/0.9.1 tqdm/4.40.2 CPython/3.6.5

File hashes

Hashes for slots-0.4.0.tar.gz
Algorithm Hash digest
SHA256 b02b7b60084b6fe2b0297a14493289c4f7ee88ae374ec518237ca6b13d39e7ae
MD5 f91161c95cf32357b2b23b01ca4041c9
BLAKE2b-256 443e20fea85cf3b7054cd90da868b50640e9dc532509567462e623ea1aeefb42

See more details on using hashes here.

File details

Details for the file slots-0.4.0-py3-none-any.whl.

File metadata

  • Download URL: slots-0.4.0-py3-none-any.whl
  • Upload date:
  • Size: 8.7 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/3.1.1 pkginfo/1.5.0.1 requests/2.22.0 setuptools/45.0.0 requests-toolbelt/0.9.1 tqdm/4.40.2 CPython/3.6.5

File hashes

Hashes for slots-0.4.0-py3-none-any.whl
Algorithm Hash digest
SHA256 2bc3e51a6223ae3984476f5d1bb272e70ca6d74fbd298fe48c7283cc1c358cc7
MD5 1ec6d4398f2a36036f452a9c77d0d43c
BLAKE2b-256 61dacfca624262fdcce7c043b5e96f06a64019d4a6581a31d5e34ee52b9d30cd

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page