Engineering / PRJ-13

Table Protocol Pong Policy — ONNX

Engineering work by Orçun Candan.

Implementation

A compact feed-forward policy published on Hugging Face for the Table Protocol Pong environment. Supervised imitation of a hand-built policy uses 20,000 generated state/action examples; 32 normalized inputs produce lane, movement and ability proposals. An ONNX Runtime Web example runs inference with the WebAssembly CPU provider.

Results & scope

Random held-out validation on 2,000 examples: 52.8% lane accuracy, 0.085 normalized movement MAE and 94.2% ability accuracy. These are scripted-teacher agreement metrics, not match win rates. No reinforcement learning or human-play training was used; cross-game generalization and improvement over the teacher have not been demonstrated.

Tools & technologies

• ONNX
• ONNX Runtime Web
• Imitation learning
• WebAssembly
• Game AI

Sources & public evidence

• Hugging Face model card — Table Protocol Pong Policy — ONNX
• Validation metrics (published revision) — Table Protocol Pong Policy — ONNX
• Browser inference example — Table Protocol Pong Policy — ONNX
• ONNX model file — Table Protocol Pong Policy — ONNX