2048-AI

A fully autonomous, self-playing 2048 game powered by an Expectimax algorithm. Built purely with HTML, CSS, and Vanilla JavaScript, this project features a modern dark-mode interface and highly optimized decision-making logic.

Features

• Expectimax Algorithm: Calculates the best possible moves by evaluating a decision tree that accounts for the random nature of new tile spawns.
• Dynamic Search Depth: Automatically scales the search depth (from 3 to 7) based on the number of empty tiles, ensuring fast early-game moves and deep calculation during tight endgame scenarios.
• Advanced Heuristics: Evaluates board states using a carefully tuned mix of smoothness, monotonicity, and empty tile counts.
• Corner Locking: Utilizes a specialized weight matrix and heavy corner penalties to ensure the highest value tile stays safely anchored in the top-left corner.
• Performance Optimized: Implements precomputed Log2 lookup tables, fast matrix math, and probabilistic pruning to evaluate thousands of moves per second without freezing the browser.
• Zero Dependencies: Built entirely from scratch using vanilla web technologies without any external libraries.
• Sleek UI/UX: A responsive, modern dark-themed interface with color-coded tiles supporting values up to 65,536.

How It Works

The AI uses the Expectimax algorithm to look ahead multiple turns. Since 2048 has an element of chance (2 or 4 tiles spawning randomly), standard Minimax isn't sufficient. Expectimax calculates the "expected value" of random events (chance nodes) rather than assuming the worst-case scenario.

To determine the strength of a board state, the AI relies on a heuristic evaluation function that scores the board based on:

1. Weight Matrix: Encourages a "snake" pattern, heavily rewarding large tiles structured in specific formations.
2. Monotonicity: Ensures tile values strictly increase or decrease along rows and columns.
3. Smoothness: Minimizes the value difference between adjacent tiles to allow for easier merging.
4. Empty Spaces: Rewards keeping the board as open as possible.

Usage

Since the project uses zero external dependencies, running it is incredibly simple:

1. Clone this repository or download the index.html file.
2. Open index.html in any modern web browser.
3. Click the Start Engine button to watch the neural net optimize and play the game automatically.