https://github.com/novoselov-ab/ai-zero

Implementation of an AlphaGo Zero paper in one C++ header file without any dependencies
https://github.com/novoselov-ab/ai-zero

alphago alphago-zero convolutional-neural-networks cpp deep-learning deep-neural-networks machine-learning mcts mcts-implementations mnist-nn reinforcement-learning self-play

Last synced: 3 days ago
JSON representation

Implementation of an AlphaGo Zero paper in one C++ header file without any dependencies

• Host: GitHub
• URL: https://github.com/novoselov-ab/ai-zero
• Owner: novoselov-ab
• License: mit
• Created: 2017-11-27T12:46:29.000Z (almost 7 years ago)
• Default Branch: master
• Last Pushed: 2018-04-18T14:11:48.000Z (over 6 years ago)
• Last Synced: 2023-12-04T20:29:15.067Z (11 months ago)
• Topics: alphago, alphago-zero, convolutional-neural-networks, cpp, deep-learning, deep-neural-networks, machine-learning, mcts, mcts-implementations, mnist-nn, reinforcement-learning, self-play
• Language: C++
• Homepage:
• Size: 11.5 MB
• Stars: 5
• Watchers: 2
• Forks: 4
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# AI-Zero

Implementation of an AlphaGo Zero paper in one C++ header file without any dependencies.

To achieve that [include/ai.h](include/ai.h) implements:

* Basic multilayer neural networks

* Multiple input and multiple output layers

* Convolution layers

* MSE and cross entropy loss

* Model serialization

* Extendable optimizer, currently: Adagrad, Adam, SGD

* Basic RL environment: game, player, replay buffer

* Monte Carlo Tree Tearch with policy NN

* Self-play, optimization, evaluation and validation workers

The idea is to have simple, minimal and easy to understand implementation. That means that performance was not a priority. Howeever the whole training process runs in parallel on 4 threads (4 workers) and optimizer also spawns some threads, it was easy to do that without sacrificing simplicity of the project.

Examples include (all single main.cpp files):

* Mnist CNN training

* Gradient checking

* Connect4 game RL training

## Building

Run `premake.bat` to generate solutions. 

It was tested only on windows. But since it's just 2 files of code for every example, it should work out of the box. There is no platform specific code.