Skip to content
This repository was archived by the owner on Jul 16, 2026. It is now read-only.

Repository files navigation

Aerythen card game rules and AI development early stages archive

UPDATE July 20206

This repository contains the artificial intelligence source code driving the Aerythen interactive demo.

Aerythen is a Dark Aero Command Line Punk universe created by David Benoy. The overarching project bridges together art, software development, an upcoming card game, and a developing series of companion novels.

Explore the Universe


Development Project Ethos

This project was about:

  • understanding algorithms by building them
  • seeing how different approaches think
  • learning by stripping problems down to their nuts and bolts

Everything is text-based, inspectable, and hackable.

Utala: KAOS 9 was the working title for the game during development.

A competitive 2-player tactical duel playable with any standard 52-card deck.

Read or Print the Game Rules (PDF) | View the Game Folder

Human gameplay screenshot

Live Player

Deployed automatically via GitHub Pages on pushes to main.

Play online

Quick Start

cd utala/kaos9
./setup.sh              # Python 3.11 venv + dependencies
./run.sh                # Play as human vs AI
make test               # Run tests

AI Research Phases

Phase 1 — Baselines (no learning) — COMPLETE

Canonical Python engine, deterministic replay, evaluation harness. Baseline agents: random legal, heuristic, Monte Carlo rollout.

Checkpoint: Is the game worth studying? PASS — Heuristic 65% vs Random, Monte Carlo 79% vs Random. Clear skill gradient with meaningful variance.

Phase 2 — Learning without frameworks — COMPLETE

Hand-built TD-linear value agent with manual gradient updates. Fixed state encoding, fixed action space, illegal actions masked by engine. Plateau ~47% vs Heuristic.

Phase 3 — Deep learning — COMPLETE

DQN with bluffing-aware 80-dim state features (39K params). Imitation learning to distill search/DQN into tiny production models.

Phase 4 — Rule evolution — COMPLETE

Variant A (v1.9): choosable dogfight order — winner picks the next contested square. Action space grows from 86 to 95. All agents retrained and validated on new rules.

Checkpoint: Is the game rich enough to require deep learning? PASS — DQN reaches 53% vs Heuristic (peak), linear models plateau at 38%.

Phase 5 — Improve, distill, ship — COMPLETE

Improved DQN to consistent >50% vs Heuristic. Distilled into tiny production model. Variant A rules and distilled agent ported to the Flutter game app.


License

This Source code is licensed under the MIT License.

Aerythen, the associated artwork, the game name "utala: kaos 9", rulebook text, and branding are © 2026 David Benoy. All rights reserved.

About

Game AI lab

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages