bayesianbandits 0.3.0

No project description provided

bayesianbandits

bayesianbandits is a Pythonic framework for building agents to maximize rewards in multi-armed bandit (MAB) problems. These agents can handle a number of MAB subproblems, such as contextual, restless, and delayed reward bandits.

Building an agent is as simple as defining arms and using the necessary decorators. For example, to create an agent for a Bernoulli bandit:

import numpy as np

from bayesianbandits import Bandit, Arm, epsilon_greedy, DirichletClassifier

def reward_func(x):
    return np.take(x, 0, axis=-1)

clf = DirichletClassifier({"yes": 1.0, "no": 1.0})
policy = epsilon_greedy(0.1)

class Agent(Bandit, learner=clf, policy=policy):
    arm1 = Arm("action 1", reward_func)
    arm2 = Arm("action 2", reward_func)

agent = Agent()

agent.pull() # receive some reward
agent.update("yes") # update with observed reward

Getting Started

Install this package from PyPI.

pip install -U bayesianbandits

Usage

Check out the documentation for examples and an API reference.