markdown

rs-ttt-rl

A Rust port of the ttt-rl project, originally written in C by antirez. This is an incomplete implementation of a Tic-Tac-Toe game using reinforcement learning with a neural network.

Project Status

This is a work-in-progress port. The following components have been implemented:

• Neural network structure with input, hidden, and output layers
• Forward pass with ReLU activation for hidden layers and softmax for output
• Weight and bias initialization using a custom pseudo-random number generator (CRand)
• Basic game state representation
• Back propagation

Missing Features (To Be Implemented)

• Training logic
• Game logic (move validation, win/draw conditions, etc.)
• Reinforcement learning integration
• Main game loop and interaction with the neural network
• Comprehensive game state tests
• Command-line interface for playing the game

Code Structure

• Neural Network (NeuralNetwork struct):

  • Supports a feedforward architecture with 18 inputs, 100 hidden neurons, and 9 outputs
  • Uses ReLU activation for hidden layers and softmax for output probabilities
  • Weights and biases initialized with random values in [-0.5, 0.5) using CRand
  • Stores intermediate activations for future backpropagation

• Game State (GameState struct):

  • Represents the Tic-Tac-Toe board as a 9-element array ('.', 'X', or 'O')
  • Tracks the current player (0 for human 'X', 1 for computer 'O')

• Pseudo-Random Generator (CRand struct):

  • Implements a linear congruential generator (LCG) compatible with C's rand()
  • Provides rand() for integers (0 to 32767) and rand_float() for floats (0 to 1)
  • Seeds with system time (seconds XOR nanoseconds) for uniqueness

• Tests:

  • Unit tests for ReLU and its derivative
  • Softmax numerical stability and correctness
  • Neural network forward pass with manual verification
  • Random weight generation within expected bounds
  • Basic game state initialization

Building and Running

Requires Rust (stable). To compile and run tests:

cargo test

The main function is currently empty, as the game logic is not yet implemented.

Dependencies

Uses only the Rust standard library (std)

Notes

• The CRand implementation aims to mimic C's rand() so that we only use the standard library
• The port prioritizes clarity and correctness over performance at this stage.