/A-B-Testing

Primary LanguageJupyter Notebook

A/B Testing with Epsilon Greedy and Thompson Sampling

Project Overview

This project explores the implementation and comparison of two popular multi-armed bandit algorithms: Epsilon Greedy and Thompson Sampling. Multi-armed bandit problems involve a trade-off between exploration (trying different arms) and exploitation (choosing the best-known arm). These algorithms are commonly used in scenarios like online advertising, recommendation systems, and more.

  1. Create a Bandit Class: Implement a class to represent the bandit arms, which are used in both Epsilon Greedy and Thompson Sampling algorithms.

  2. Create EpsilonGreedy() and ThompsonSampling() Classes and Methods: Implement classes and methods for the Epsilon Greedy and Thompson Sampling algorithms, both of which are inherited from the Bandit class.

    • Epsilon Greedy:

      • Implement an Epsilon Greedy class that decays epsilon by 1/t.
      • Design the experiment using the Epsilon Greedy algorithm.

    • Thompson Sampling:

      • Implement a Thompson Sampling class with known precision.
      • Design the experiment using the Thompson Sampling algorithm.

  3. Report:

    • Visualize the Learning Process: Implement a method plot1() to visualize the learning process for each algorithm.
    • Visualize Cumulative Rewards: Compare and visualize cumulative rewards from Epsilon Greedy and Thompson Sampling.
    • Store the Rewards in a CSV File: Create CSV files that store experiment data, including Bandit, Reward, and Algorithm information.
    • Print Cumulative Reward: Display the cumulative reward achieved by the algorithms.
    • Print Cumulative Regret: Display the cumulative regret of each the algorithms.

Structure

  • bandit.py: Defines the abstract class for bandit arms, Epsilon Greedy and Thompson Sampling algorithm implementations and functions for visualizing experiment results.
  • Report.ipynb: A report of experiment results.