Repository navigation
Adaptive Search
APORIA's campaign driver spends a finite execution budget over the declared parameter space. The driver is shared by all three benchmark strategies; the strategies differ in how they choose points. This makes their comparison primarily a comparison of sampling choices, while the evidence and Atlas machinery stay common.
- Random samples uniformly across the declared domain.
-
Stratified uses a Halton low-discrepancy sequence for broad coverage. The command-line strategy name is
stratified;haltonis accepted as an alias. - Adaptive chooses among informative acquisition families while retaining a floor of random exploration.
Adaptive's available families are coverage, boundary, uncertainty, sensitivity, and contradiction. Coverage continues broad sampling; boundary targets cells carrying evidence; uncertainty favours low-sample or risk-variable cells; sensitivity re-probes an axis implicated by perturbations; contradiction targets locations where channels disagree. The shared family type also includes random, which adaptive can choose during exploration.
Each acquisition family tracks its number of evaluations and an exponential average of observed information gain per evaluation. A UCB-style score favours families that have not been tried recently as well as those that have performed well. The default policy uses an exploration probability of 0.08 and average update rate of 0.1. Early gains can be noisy, so adaptive retains random exploration rather than committing permanently to the first apparent winner.
An acquisition receives credit for effects such as a newly suspicious cell, a tightened boundary band, or increased resolved coverage. The point selection and these credits are recorded as campaign decisions. These are implementation choices and are themselves open to empirical evaluation; see Experiments & Results.
The configured budget counts executions, including extra runs used for probes. A perturbation pair, symmetry swap, f32 comparison, or independent-reference comparison therefore consumes budget when the executor supports and the configured rate enables it. Symmetry swaps are only run when the swapped coordinates remain valid for the paired domains. The campaign avoids buying precision or differential probes from an executor that cannot provide a genuinely distinct path.
Instruction steps are tracked separately from the number of evaluations. They are the deterministic work measure used in benchmark comparisons; elapsed wall-clock time is recorded but is affected by machine and runtime conditions. See Benchmarking.
Adaptivity cannot reliably find a region the evaluations never touch, and search-point selection cannot overcome the Atlas's maximum depth or axis-aligned representation. The current measured results are mixed: adaptive localised the narrow rlc_resonance band in one seed at the largest tested budget while baselines did not localise it; simple coverage did better on other entries. See Limitations.