---
title: "Custom strategies"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Custom strategies}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r setup, include = FALSE}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)
```

ppforest2 trains trees by composing six pluggable strategies:

| Strategy | Purpose | Built-in |
|----------|---------|----------|
| **PP** (projection pursuit) | Find the projection that best separates groups | `pp_pda()` — Penalized Discriminant Analysis |
| **Vars** (variable selection) | Select which variables are available at each split | `vars_uniform()`, `vars_all()` |
| **Threshold** (split cutpoint) | Compute the split cutpoint in projected space | `cutpoint_mean_of_means()` |
| **Stop** (stopping rule) | Decide when to stop growing | `stop_pure_node()` |
| **Binarize** (binarization) | Reduce multiclass to binary at each node | `binarize_largest_gap()` |
| **Grouping** (group partition) | Route observations to children | `grouping_by_label()` |

You can add new strategies without modifying the core tree-building logic. This vignette walks through the process.

## How strategies work

Each strategy is an R list with a `name` field that identifies it, a `display_name` for summaries, and any parameters the strategy needs. The `name` must match a C++ strategy registered under the same name.

```{r inspect}
library(ppforest2)

pp_pda(0.5)
vars_uniform(n_vars = 2)
cutpoint_mean_of_means()
```

When you call `pptr()` or `pprf()`, the strategy lists are passed to C++, where the `name` field dispatches to the corresponding C++ implementation. The actual computation (optimization, variable selection, cutpoint) happens entirely in C++.

## Adding a new strategy

Adding a strategy requires work on both sides:

1. **C++**: Implement the strategy class (the computation).
2. **R**: Write a constructor function (the user-facing API).

### Step 1: C++ implementation

Each strategy family has a base class with pure virtual methods. Your new strategy inherits from the appropriate base and implements them.

For example, a new projection pursuit strategy needs to implement `optimize()` (find the best projection):

```cpp
// File: core/src/models/strategies/pp/MyMethod.hpp
#pragma once
#include "models/strategies/pp/ProjectionPursuit.hpp"
#include "models/strategies/Strategy.hpp"
#include "utils/JsonValidation.hpp"

namespace ppforest2::pp {

struct MyMethod : public ProjectionPursuit {
  explicit MyMethod(float alpha) : alpha_(alpha) {}

  std::string display_name() const override { return "My method"; }

  Result optimize(
    const types::FeatureMatrix&  x,
    const stats::GroupPartition& group_spec) const override {
    // Find the optimal projector for the data.
    // Return Result{ projector_vector, index_value }.
    ...
  }

  nlohmann::json to_json() const override {
    return {{"name", "my_method"}, {"alpha", alpha_}};
  }

  static ProjectionPursuit::Ptr from_json(const nlohmann::json& j) {
    JsonReader{j, "my_method"}.only_keys({"name", "alpha"});
    return my_method(j.at("alpha").get<float>());
  }

  PPFOREST2_REGISTER_STRATEGY(ProjectionPursuit, "my_method")

private:
  const float alpha_;
};

inline ProjectionPursuit::Ptr my_method(float alpha) {
  return std::make_shared<MyMethod>(alpha);
}

}  // namespace ppforest2::pp
```

The key pieces:

- **`to_json()`** serializes the strategy name and parameters. This is used for model persistence.
- **`from_json()`** deserializes from JSON and validates that no unexpected keys are present.
- **`PPFOREST2_REGISTER_STRATEGY`** registers the factory so JSON deserialization finds it automatically.
- **`display_name()`** returns a human-readable label for summaries.
- **Factory function** (`my_method()`) is a convenience wrapper.

The same pattern applies to variable selection strategies (`select()`), cutpoint strategies (`cutpoint()`), and the other strategy families. See the C++ documentation for complete interface definitions and examples.

After writing the `.cpp` file, add it to `core/src/models/CMakeLists.txt`.

### Step 2: R constructor

Write an R function that creates a strategy list. The `name` field must match the C++ registration name exactly.

```{r, eval = FALSE}
#' My custom projection pursuit strategy.
#'
#' @param alpha A tuning parameter.
#' @return A \code{pp_strategy} object.
#' @export
pp_my_method <- function(alpha = 1.0) {
  if (!is.numeric(alpha) || length(alpha) != 1)
    stop("`alpha` must be a single number.")

  structure(
    list(name = "my_method", display_name = "My method", alpha = alpha),
    class = "pp_strategy"
  )
}
```

The constructor should:

- **Validate parameters** before they reach C++. Catching errors early with clear messages is better than a C++ exception.
- **Set the S3 class** to `pp_strategy`, `vars_strategy`, `cutpoint_strategy`, `stop_strategy`, `binarize_strategy`, or `grouping_strategy`. This is checked by `resolve_strategies()`.
- **Include `display_name`** for readable output in `summary()`.
- **Use the same parameter names** as `to_json()` in C++. The R list is converted to JSON and passed to `from_json()` on the C++ side.

### Step 3: Use it

Once both sides are in place, the new strategy works like any built-in:

```{r, eval = FALSE}
# Single tree
tree <- pptr(Species ~ ., data = iris, pp = pp_my_method(alpha = 0.5))

# Forest
forest <- pprf(Species ~ ., data = iris, pp = pp_my_method(alpha = 0.5), vars = vars_uniform(n_vars = 2))

# Summary shows the strategy
summary(tree)
```

The strategy is also available from the CLI:

```bash
ppforest2 train -d iris.csv --pp my_method:alpha=0.5
```

And models trained with the new strategy can be saved and loaded as usual --- the JSON registry handles serialization automatically.

## Strategy families reference

### PP: Projection pursuit

Controls how the tree finds the best linear combination of variables at each node.

```
index(x, group_spec, projector) -> scalar
optimize(x, group_spec) -> Result{projector, index}
```

`optimize()` is the main method. It receives the data matrix and group partition and returns the best projection vector. `index()` evaluates a given projection (used for variable importance calculations).

### Vars: Variable selection

Controls which variables are available to projection pursuit at each split. This is what makes random forests "random".

```
select(x, group_spec, rng) -> Result{selected_indices, original_cols}
```

The returned `VariableSelection::Result` tracks which columns were selected so the reduced-space projector can be expanded back to the full feature space.

### Threshold: Split cutpoint

Controls where the split cutpoint is placed in the projected space.

```
cutpoint(group_1, group_2, projector) -> scalar
```

Receives the two groups (already partitioned by projection pursuit) and the projection vector. Returns the cutpoint value.

### Stop: Stopping rule

Controls when to stop growing the tree.

```
should_stop(group_partition, depth) -> bool
```

### Binarize: Binarization

Controls how multiclass nodes (>2 groups) are reduced to a binary problem.

```
regroup(projected_x, group_partition) -> Result
```

### Grouping: Group partition

Controls how observations are routed to children after a split.

```
split(partition, lower_group, upper_group) -> Result
```

## Checklist

1. Create `core/src/models/strategies/<family>/MyStrategy.hpp` (and `.cpp` if needed).
2. Inherit from the appropriate base class (`ProjectionPursuit`, `VariableSelection`, `Cutpoint`, `StopRule`, `Binarization`, or `Grouping`).
3. Implement the pure virtual methods.
4. Implement `to_json()` with a `"name"` field.
5. Implement `display_name()` for human-readable summaries.
6. Add `static Ptr from_json()` with key validation.
7. Add `PPFOREST2_REGISTER_STRATEGY(Base, "name")`.
8. Add a factory function in the strategy's namespace.
9. Add the `.cpp` to `core/src/models/CMakeLists.txt`.
10. Write tests in `MyStrategy.test.cpp` (JSON round-trip + functional).
11. Write the R constructor function with validation, `display_name`, and the correct S3 class.
12. Export and document the R function.