Bandits Playground

Taking inspiration from Sutton & Barto this repo is a place to experiment with bandits algorithms.

Included Bandit Algorithms

See policies.py

• Random
• Greedy
• Epsilon Greedy
• Softmax
• UCB1

Requirements

• Python 3.10
  • poetry

Installation

1. Use Poetry to install dependencies

    poetry install
   

Usage

Run the experimenter from main.py with command-line arguments

poetry run main.py {options}

For example

poetry run main.py --nb_bandits=100 --bandit_type=gaussian --steps=500

Options

Flag	Parameters	Description	Required	Default Value
nb_bandits	int	The number of bandit arms	N	10
bandit_type	{bernoulli, gaussian}	How the bandit distributes rewards	N	bernoulli
steps	int	How many steps to train for	N	1000
trials	int	How many to times to repeat the experiment	N	5