← Back to the chapter · Previous detail · Next detail →
Technical detail · ACE and Shared Context Sensing
Bound planning work and measure the approximation
A practical bounded heuristic proceeds in six steps:
- Prune invalid evidence. Remove stale cache entries, forbidden sensors and rules below threshold.
- Estimate marginal value. Estimate how much probability mass each test moves to a valid terminal answer.
- Score each test. A useful greedy score is . Charge a shared wake only once when tests run in the same active window.
- Expand a small beam. Keep the best partial trees to retain alternatives when two tests interact.
- Attach an authoritative fallback. End unresolved or low-confidence branches with a known sensor path or an explicit “unknown.”
- Measure regret. Replay traces against exact search on small cases; record differences in energy and answer quality.
The request path remains cache, rule, proxy, then target sensing. Check cached age against the request’s maximum age and confidence against its minimum. Rule inference also needs fresh, permitted antecedents; a proxy plan must meet expected confidence and return the inferred target attribute, not just the raw proxy observation. Return evidence with each result and record why earlier options failed.
Re-plan when energy costs, cache-hit rates or correlations drift. A sensing planner is an adaptive policy maintained with evidence, rather than a one-time sensor ordering.