MEOK partial-observation process predictor v1
A three-member MLP output ensemble trained on fresh synthetic procurement-process episodes without the simulator's full hidden profile proportions.
Evaluation
The held-out test contains 560 rows from 70 episode IDs (1250–1319). The model is selected on validation only.
- Partial-observation best single MLP + simplex/expiry rules: MAE 0.0234502.
- Exported three-model output ensemble + the same rules: MAE 0.0188549.
- Persistence: MAE 0.0613514.
- Original v0 privileged-input model + rules: MAE 0.0016888. This comparator sees three additional hidden profile proportions and is not a like-for-like partial-observation comparison.
The output ensemble averages predictions; it does not merge model weights. The specialist policy router was tested and rejected because it performed worse.
Input and output
Twelve inputs: previous four-state aggregate, previous-previous four-state aggregate, revocation indicator, ticks until permit expiry divided by eight, policy-mode indicator and tick divided by eight. Four outputs: next aggregate state proportions.
Use predict.py for dependency-free inference. It returns raw and rule-constrained outputs with an explicit NONE authority marker.
Limitations
This is a deterministic synthetic procurement-process simulator, not a physical world model, real market model, general-purpose agent or autonomous action system. The expiry rule accounts for zero error on expired rows; this is a known rule, not learned behaviour. Worst held-out component error remains above 0.52. Do not use for real decisions or claim real-world capability.
Source and reproducibility: https://github.com/CSOAI-ORG/councilof-ai/pull/4285
Browser demo and reproduction bundle
- Interactive browser demo: https://csoai-meok-oowm-hybrid-lab.static.hf.space/partial-observation-v1/index.html
- Reproducible source, data splits, weights and tests: MEOK-PARTIAL-OBSERVATION-V1.zip
- Canonical source and CI: https://github.com/CSOAI-ORG/councilof-ai/pull/4285