🧪 勝負預測實驗室
產生對戰，用前 70% 建立玩家輪廓，訓練模型預測後 30% — 全程在瀏覽器中完成。
💡 為什麼不需要 GPU？
- Logistic regression has 15 parameters; even the deep MLP has ~2.2k — 100k matches of data is just a few MB.
- One epoch is millions of multiply-adds — milliseconds to seconds in a modern JS engine (see measured times above).
- At this scale WebGL/WebGPU kernel-dispatch overhead actually loses to plain CPU TypeScript.
- A real community produces hundreds to thousands of games — full retraining after every match stays under 100ms.