# Classifier optimization

Classifier optimization is an optional post-classify stage for neural readout
methods. The canonical import path is ``arcade.classifieroptimization``. The
module ``arcade.optimization`` is a compatibility shim that re-exports the
same public API.

Enable the stage through YAML as shown in
{doc}`../how_to/enable_optimization`.

## Public surface

| Symbol | Role |
|--------|------|
| ``BaseOptimizer`` | Optimizer contract |
| ``OptimizationResult`` | Structured result object |
| ``register_optimizer``, ``get_optimizer``, ``list_optimizers`` | Registry helpers |
| ``PruningOptimizer`` | Weight pruning under an error budget |
| ``DistillationOptimizer`` | Teacher to student distillation |
| ``NASOptimizer`` | Architecture search |
| ``HyperparameterTuner`` | Optuna or grid search via ``tune`` and ``maybe_tune`` |
| ``run_optimization_stage`` | Pipeline stage entry point |
| ``pareto_frontier``, ``select_best`` | Multi-objective selection helpers |

Quantization uses ``arcade.nn.quantization.quantize_model`` from the stage
(no ``QuantizationOptimizer`` class).

## Config models

The top-level ``optimization`` section maps to ``OptimizationConfig`` in
``arcade.config``:

| Field | Default | Notes |
|-------|---------|-------|
| ``enabled`` | ``false`` | Master switch |
| ``target_classifiers`` | empty list | Empty means all eligible neural methods |
| ``error_budget`` | ``0.01`` | Allowed accuracy drop |
| ``pruning``, ``distillation``, ``quantization``, ``nas`` | nested configs | Per-technique flags |

Hyperparameter search uses the sibling ``tuning`` block as ``TuningConfig``.
It is not nested under ``optimization``.

## Example

```yaml
optimization:
  enabled: true
  target_classifiers: [fnn]
  error_budget: 0.01
  pruning:
    enabled: true

tuning:
  enabled: false
```

```python
from arcade.classifieroptimization import (
    list_optimizers,
    pareto_frontier,
    run_optimization_stage,
)

print(list_optimizers())
```

Autosummary pages live under {doc}`api/index`.