connect-tycoon
Korean RL tycoon game training a company decision model that answers situations with choice, probabilities and confidence.
Overview
CONNECT TYCOON is an open-source reinforcement-learning game from the CONNECT AI LAB education brand. A model runs a company from the top floor of a building; each decision flows down to the floors and comes back as money, customer satisfaction or incidents. Pressing train lets the model run the company hundreds of times in the simulator, learning first by imitating a rule-based boss and then by PPO, and it asks a human when it is not confident.
Each company decision is one request with a state and criteria, answered with choice, probabilities and confidence. The README says this matches the shape of `typesafe/jev-1.13` on OpenRouter, so any model answering in that format can be trained and tested in the simulator. It exposes a Gymnasium environment with action masks, Colab notebooks to upload a trained model or move it to Qwen, and a first result model on Hugging Face. Training runs on a laptop CPU.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Given this company situation, which action should be taken?(inferred) | — |
How it works
- Simulator presents a business situation and options
- The decision model returns a choice with probabilities and confidence
- Low confidence escalates to a human
- Outcomes (money, satisfaction, incidents) feed PPO training
Strengths
- Runs on CPU with no GPU required
- Human-in-the-loop on low confidence
- Jev-compatible request shape so other models can be compared
- Gymnasium environment for developers
Caveats
- Educational project; README is in Korean
- Results are from a simulator, not real business data
Getting started
git clone https://github.com/wonseokjung/connect-tycoon, create a venv, pip install -e ".[game]" and run connect-tycoon play.At a glance
- Type
- Project
- Status
- prototype
- Language
- Python
- Runtime
- desktop game + Gymnasium
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
시뮬레이터로 기업 의사결정 모델을 학습시키는 오픈소스 강화학습 게임 · JEV · PPO · Human-in-the-loop.
wh000wh000/awesome-jev-live
Frequently asked questions
Does it use the Jev API?
It uses the same request and answer shape as typesafe/jev-1.13, so Jev-format models can be trained and tested in it.
Do I need a GPU?
No, the README says training runs on a laptop CPU.
What does the model decide?
Business actions for a simulated company, one decision per step.
Related projects
Fully local grid tactics game where the open NanoJev model picks each move, attack, heal or dash from real candidate actions.
ProjectresearchPythonToy archipelago air-traffic sim where each plane asks Jev every 1.6 s whether to divert, declare an emergency, give way or land.
DemoprototypeTypeScriptInteractive Jev experiments, from a support-message decision lab to 3D self-driving demos with steer, speed and overtake decisions.
AppprototypeJev-compatible open System One engine with a Claude Code-style terminal REPL showing live probabilities, plus an experiment harness.
ProjectprototypePython33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
ProjectresearchPythonOpen 4B System I driving model that scores behaviours like cruise, stop, yield and brake from three camera frames each decision.
ProjectresearchPython