A python implementation of Exponential weighting for Exploration and Exploitation with Experts. Based on this blog post.

This algorithm is useful for non-stochastic Contextual Multi Armed Bandits.

Table of Contents

• Installation
• Usage

If you just need to use exp4, you can just run:

$ pip install exp4

For developers, note that this project uses the poetry python package/dependency management tool. Please familarize yourself with it and then run:

$ poetry install

exp4 is centered around the exp4.exp4 function which creates a co-routine for selecting arms given expert advice.

The protocol is as follows:

1. The expert constructs an expert advice matrix.
   • Each row contains the corresponding experts advice vector.
   • The advice vector provides probabilities for each arm.
2. The expert sends a tuple of loss and advice.
   • The loss corresponds to the previous round.
   • The first round's loss is ignored.
   • The advice correspond to the current round.

An example is given below.

from exp4 import exp4

player = exp4()

loss = None           # Will be ignored.
advice = [
    [1/3, 1/3, 1/3],  # Expert 1 
    [2/3, 1/3, 0],    # Expert 2
]
arm = player.send((loss, advice))
assert arm in range(3)

loss = 1 / (1 + arm)  # Arbitrary loss assigned to arm.
advice = [
    [0, 0, 1],        # Expert 1
    [0, 0, 1],        # Expert 2
]
arm = player.send((loss, advice))
assert arm == 2