policy

No-reset classification policies for continuous detector traces.

In no-reset benchmarking the detector runs continuously without restarting after each alarm. A policy decides which thresholded points on that continuous detection function count as true positives, false positives, or false negatives around a true change point (or inside a no-change regime).

Policies do not modify the underlying trace. They translate an infinite-threshold run into a threshold-marked NoResetDetectionTrace that downstream metrics (precision, recall, F1, ARL, …) can consume.

Public API

  • NoResetPolicy – protocol that all policies implement. The core method is apply(run, threshold) -> SingleRun[NoResetDetectionTrace, ProviderT].

  • BisegmentPolicyBase – abstract base for policies that operate on single-change-point bisegment providers. Subclasses define how to select detections in the false region (before the change point) and the true region (at or after the change point).

  • PointBasedPolicy – treats every threshold-satisfying point as a candidate detection. In the true region only the first such point is kept.

  • EventBasedPolicy – focuses on threshold-crossing events (the first point at or above threshold after being below). In the true region only the first event is kept.

  • MixedPolicy – uses event-based selection in the false region and point-based selection in the true region.

  • NoChangePolicy – for no-change (ARL) providers. Keeps at most the first point that satisfies the threshold.

See the docstrings in base, bisegment, and nochange submodules for implementation details of each class.

Examples

Examples

Create a policy and apply it to a single run:

>>> from pysatl_cpd.benchmark.online.noreset.metrics.policy import (
...     MixedPolicy,
...     PointBasedPolicy,
...     EventBasedPolicy,
...     NoChangePolicy,
... )
>>> point_policy = PointBasedPolicy(max_delay=15, strict=False)
>>> event_policy = EventBasedPolicy(max_delay=15, strict=False)
>>> mixed_policy = MixedPolicy(max_delay=15, strict=False)
>>> nochange_policy = NoChangePolicy(strict=False)

Policies are typically consumed through OnlineNoResetBenchmark rather than called directly. The benchmark constructs the appropriate policy from a NoResetPolicyKind:

>>> from pysatl_cpd.benchmark.online.noreset import (
...     NoResetPolicyKind,
...     OnlineNoResetBenchmark,
... )
>>> # benchmark = OnlineNoResetBenchmark(
... #     dataset=dataset,
... #     registry=registry,
... #     max_delay=15,
... #     global_policy=NoResetPolicyKind.MIXED,
... #     error_margin=(0, 15),
... #     policy_strict=False,
... # )

Notes

Notes

  • All bisegment policies (PointBasedPolicy, EventBasedPolicy, MixedPolicy) require the run provider to be a bisegment with exactly one true change point.

  • NoChangePolicy requires the run provider to be a no-change provider (provider_type == ProviderType.NO_CHANGE).

  • The strict parameter controls whether threshold comparison uses strict (>) or non-strict (>=) inequality.

  • max_delay must be non-negative and defines the size of the true region as [cp, min(cp + max_delay, length - 1)].

  • Policies are stateless after construction and can be reused across multiple runs and thresholds.