zeromodel

Spatial optimizer

The spatial optimizer is the first core bridge from manual dense views to spatial calculus.

A ViewProfile says which metrics to turn up or down. A SpatialOptimizer learns a non-negative metric-weight profile for one explicit objective:

concentrate high-signal mass in the top-left inspection region.

It does not prove that the resulting view is semantically correct, universally optimal, or better for every task. It optimizes a deterministic geometric objective over scored data.

Example

from zeromodel import ScoreTable, SpatialOptimizer, build_optimized_view, optimize_view_profile

source = ScoreTable(
    values=[
        [0.10, 0.50, 0.20],
        [0.95, 0.50, 0.25],
        [0.90, 0.50, 0.15],
        [0.05, 0.50, 0.20],
    ],
    row_ids=["background", "target_a", "target_b", "flat"],
    metric_ids=["target", "constant", "weak"],
)

optimizer = SpatialOptimizer(Kc=2, Kr=2, alpha=0.95, max_iters=40)
result = optimize_view_profile(source, name="optimized-target", optimizer=optimizer)
view = build_optimized_view(source, name="optimized-target", optimizer=optimizer)

print(result.metric_weights)
print(result.baseline_mass, result.optimized_mass)
print(view.cell(0, 0).row_id, view.cell(0, 0).metric_id)

What is optimized

For a candidate metric-weight vector, the optimizer:

1. normalizes each metric to [0, 1],
2. sorts columns by learned weight,
3. sorts rows by weighted intensity in the top Kc columns,
4. computes spatially decayed mass in the top Kr × Kc block,
5. uses deterministic coordinate ascent to improve that score.

Why it matters

This turns the old spatial-calculus idea into a repo-backed primitive:

same dense source table
→ learned metric weights
→ ViewProfile
→ optimized VPM view
→ same source digest and source mapping

The next step after this is not to claim universal optimization. The next step is benchmark fixtures that compare manual views, optimized views, and task-specific outcomes.

This site is open source. Improve this page.