Multi-Armed Bandit Web Dashboard 🎰

This is a purely client-side React Single Page Application (SPA) designed to mathematically simulate and visually analyze various Multi-Armed Bandit (MAB) reinforcement learning algorithms.

Try it live here: 👉 Live Dashboard on GitHub Pages

Overview

This application performs a Monte-Carlo simulation involving 10,000 steps averaged over 100 independent trials. The environment consists of 3 Bandits with the following True Means:

• Bandit A: 0.80 (Optimal)
• Bandit B: 0.70
• Bandit C: 0.50

The heavy numerical simulation is offloaded seamlessly to a background JavaScript Web Worker, preventing the UI thread from freezing and allowing for instant reactivity even processing billions of randomized calculations.

Implemented Algorithms

1. A/B Testing (Explore-then-Exploit)
2. Optimistic Initial Values
3. $\epsilon$-Greedy
4. Softmax (Boltzmann Distribution)
5. Upper Confidence Bound (UCB)
6. Thompson Sampling (Bayesian)

Dashboard Features

1. General Comparison View

Compares all 6 algorithms iteratively across the full 10,000 budget, tracing their Cumulative Average Return and their % Optimal Action Selection to demonstrate how intelligent exploration methods (like UCB and Thompson) absolutely crush static A/B testing methods.

2. A/B Testing Deep Dive

An exclusive dashboard dedicated to analyzing the classic Explore: $2,000 / Exploit: $8,000 strategy exactly matching standard statistical requirements:

• Chart 1: Line chart visualizing the volatility and stabilization of estimated empirical means prior to exploitation.
• Chart 2: A standard deviation ($\pm 1 \text{ SD}$) bounded, cumulative running average curve clearly demarcating the Phase 1 vs Phase 2 borders.
• Chart 3: A discrete Grouped Bar comparison plotting exactly what the machine believed at Step 2000 against actual mathematical reality.

Local Development

To run this project locally:

1. Clone this repository.
2. Install dependencies:

   npm install

3. Start the Vite dev server:

   npm run dev

4. Access it in your browser natively at http://localhost:5173.

---

Built with React, Vite, standard Chart.js / react-chartjs-2, and modern CSS.