A full-stack Gomoku (five-in-a-row) game with two AI engines, built as a two-person École 42 project: a C++ Minimax engine (alpha-beta with iterative deepening, a transposition table and a PVS hard mode) and a Python AlphaZero engine (self-play reinforcement learning with MCTS + ResNet).
| Layer | Tech |
|---|---|
| Frontend | Nuxt 3, Vue 3, TypeScript |
| Minimax engine | C++, libwebsockets, Zobrist hashing, PVS with transposition table |
| AlphaZero engine | Python, PyTorch, MCTS (C++ native extension), ONNX inference |
| Training infra | Ray cluster on GCP, self-play + arena evaluation pipeline |
| Deployment | Docker containers: AlphaZero on a Google Cloud VM, Minimax on an Oracle Cloud Arm VM (since Sep 2026); Cloudflare Workers routing |
A two-person 42 project by Sungyong Cho and Woolim Park.
- Sungyong Cho: the AlphaZero training pipeline, the C++ Minimax search, the C++ (libwebsockets) and FastAPI WebSocket servers, the deployment, the documentation pages and four interactive diagram components. Designed the hard-mode evaluation (pattern set, precomputed pattern-score lookup tables, incremental re-evaluation and move ordering), most of which Woolim implemented.
- Woolim Park: the game UI and most of the hard-mode evaluation implementation.
- Together: the WebSocket connection handling, since client and server had to change together.
Requires Docker and Docker Compose v2+.
git clone git@github.com:sungyongcho/gomoku.git
cd gomoku
cp .env.example .env
source alias.sh
# full stack (frontend + minimax + alphazero)
dev-up
# frontend + minimax only (no AlphaZero)
docker compose -f docker-compose.yml up front minimaxOpen http://localhost:${LOCAL_FRONT}/gomoku (default port: 3000).
Local documentation: http://localhost:${LOCAL_FRONT}/gomoku/docs/about-gomoku/intro
See .env.example for all configurable ports and variables.
alias.sh— Docker Compose shortcuts (dev-up,dev-down,dev-up-debug, etc.)Makefile— Original École 42 project submission entry point (front + minimax only)front/— Nuxt 3 frontend with interactive documentation and diagramsminimax/— C++ minimax engine (WebSocket server)alphazero/— AlphaZero training pipeline, inference server, and C++ MCTS extension
- Double-three detection uses logic derived from the Renju open source reference implementation. Refer to the upstream page for license details and attribution.
- The AlphaZero engine started in May 2025 from the AlphaZeroFromScratch tutorial and continued in private repositories (Aug 2025 - Feb 2026) before the rewritten version was imported here.