Scheduler interface

[andrewdalpino]

Add Scheduler interface for scheduler optimizers. Advance every batch instead of every parameter update.

A 3-Dense MLP with biases (6 Parameters per batch) now advances t by exactly 1 per batch instead of 6 — so the user-configured steps / length / decay parameters behave at the batch scale the docs promise, instead of at the parameter-update scale they were silently ticking at before.

[Copilot]

🔵 Needs a closer look

It changes core training-loop semantics for learning-rate schedules and should be validated with real training runs/benchmarks in addition to the included unit tests.

Pull request overview

This PR introduces a Scheduler optimizer contract so global learning-rate schedules advance once per batch (forward+backward pass) rather than once per parameter update, aligning runtime behavior with the documented meaning of “steps”/schedule length.

Changes:

• Added Scheduler interface with a tick() method, implemented by StepDecay and Cyclical.
• Updated FeedForward::roundtrip() to call tick() once per batch for scheduler optimizers.
• Added/updated PHPUnit coverage to ensure schedulers tick per batch and updated optimizer docs to define “step” as “batch”.

File summaries

File	Description
tests/NeuralNet/Optimizers/StepDecayTest.php	Adds a regression test ensuring StepDecay advances its schedule once per batch, not per parameter.
tests/NeuralNet/Optimizers/SchedulerTest.php	Verifies which optimizers implement the new Scheduler contract.
tests/NeuralNet/Optimizers/CyclicalTest.php	Adds a regression test ensuring Cyclical advances its schedule once per batch, not per parameter.
src/NeuralNet/Optimizers/StepDecay.php	Implements Scheduler and moves schedule advancement into tick().
src/NeuralNet/Optimizers/Scheduler.php	Introduces the Scheduler interface defining tick() for batch-level schedule advancement.
src/NeuralNet/Optimizers/Cyclical.php	Implements Scheduler and moves schedule advancement into tick().
src/NeuralNet/FeedForward.php	Advances scheduler optimizers exactly once per roundtrip() (per batch).
docs/neural-network/optimizers/step-decay.md	Clarifies that a “step” corresponds to one batch and updates parameter wording accordingly.
docs/neural-network/optimizers/cyclical.md	Clarifies batch semantics, aligns parameter naming (length) and updates the example signature.

Review details

• Files reviewed: 9/9 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.