solverforge_solver/phase/localsearch/

phase.rs

//! Local search phase implementation.

use std::fmt::Debug;
use std::marker::PhantomData;
use std::time::Instant;

use rand::rngs::SmallRng;
use rand::SeedableRng;
use solverforge_core::domain::PlanningSolution;
use solverforge_scoring::{RecordingScoreDirector, ScoreDirector};
use tracing::{debug, info, trace};

use crate::heuristic::r#move::{Move, MoveArena};
use crate::heuristic::selector::MoveSelector;
use crate::phase::localsearch::{Acceptor, LocalSearchForager};
use crate::phase::Phase;
use crate::scope::BestSolutionCallback;
use crate::scope::{PhaseScope, SolverScope, StepScope};

/// Local search phase that improves an existing solution.
///
/// This phase iteratively:
/// 1. Generates candidate moves into an arena
/// 2. Evaluates each move by index
/// 3. Accepts/rejects based on the acceptor
/// 4. Takes ownership of the best accepted move from arena
///
/// # Type Parameters
/// * `S` - The planning solution type
/// * `M` - The move type
/// * `MS` - The move selector type
/// * `A` - The acceptor type
/// * `Fo` - The forager type
///
/// # Zero-Clone Design
///
/// Uses index-based foraging. The forager stores `(usize, Score)` pairs.
/// When a move is selected, ownership transfers via `arena.take(index)`.
/// Moves are NEVER cloned.
pub struct LocalSearchPhase<S, M, MS, A, Fo>
where
    S: PlanningSolution,
    M: Move<S>,
    MS: MoveSelector<S, M>,
    A: Acceptor<S>,
    Fo: LocalSearchForager<S, M>,
{
    move_selector: MS,
    acceptor: A,
    forager: Fo,
    arena: MoveArena<M>,
    step_limit: Option<u64>,
    _phantom: PhantomData<fn() -> (S, M)>,
}

impl<S, M, MS, A, Fo> LocalSearchPhase<S, M, MS, A, Fo>
where
    S: PlanningSolution,
    M: Move<S> + 'static,
    MS: MoveSelector<S, M>,
    A: Acceptor<S>,
    Fo: LocalSearchForager<S, M>,
{
    /// Creates a new local search phase.
    pub fn new(move_selector: MS, acceptor: A, forager: Fo, step_limit: Option<u64>) -> Self {
        Self {
            move_selector,
            acceptor,
            forager,
            arena: MoveArena::new(),
            step_limit,
            _phantom: PhantomData,
        }
    }
}

impl<S, M, MS, A, Fo> Debug for LocalSearchPhase<S, M, MS, A, Fo>
where
    S: PlanningSolution,
    M: Move<S>,
    MS: MoveSelector<S, M> + Debug,
    A: Acceptor<S> + Debug,
    Fo: LocalSearchForager<S, M> + Debug,
{
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("LocalSearchPhase")
            .field("move_selector", &self.move_selector)
            .field("acceptor", &self.acceptor)
            .field("forager", &self.forager)
            .field("arena", &self.arena)
            .field("step_limit", &self.step_limit)
            .finish()
    }
}

impl<S, D, BestCb, M, MS, A, Fo> Phase<S, D, BestCb> for LocalSearchPhase<S, M, MS, A, Fo>
where
    S: PlanningSolution,
    D: ScoreDirector<S>,
    BestCb: BestSolutionCallback<S>,
    M: Move<S>,
    MS: MoveSelector<S, M>,
    A: Acceptor<S>,
    Fo: LocalSearchForager<S, M>,
{
    fn solve(&mut self, solver_scope: &mut SolverScope<S, D, BestCb>) {
        let mut phase_scope = PhaseScope::new(solver_scope, 0);
        let phase_index = phase_scope.phase_index();

        // Calculate initial score
        let mut last_step_score = phase_scope.calculate_score();

        info!(
            event = "phase_start",
            phase = "Local Search",
            phase_index = phase_index,
        );

        // Notify acceptor of phase start
        self.acceptor.phase_started(&last_step_score);

        let start_time = Instant::now();
        let mut local_moves_evaluated: u64 = 0;
        let mut last_progress_time = Instant::now();
        let mut last_progress_moves: u64 = 0;
        let mut rng = SmallRng::from_os_rng();

        loop {
            // Check early termination
            if phase_scope.solver_scope().should_terminate() {
                break;
            }

            // Check step limit
            if let Some(limit) = self.step_limit {
                if phase_scope.step_count() >= limit {
                    break;
                }
            }

            let mut step_scope = StepScope::new(&mut phase_scope);

            // Reset forager and acceptor for this step.
            // Pass best and last-step scores so foragers that implement
            // pick-early-on-improvement strategies know their reference targets.
            let best_score = step_scope
                .phase_scope()
                .solver_scope()
                .best_score()
                .copied()
                .unwrap_or(last_step_score);
            self.forager.step_started(best_score, last_step_score);
            self.acceptor.step_started();

            // Regenerate moves every step so the move space reflects the
            // current solution topology. This is essential for list variables
            // (VRP-style) where positions change after each accepted move.
            self.arena.reset();

            if self.move_selector.is_never_ending() {
                // Streaming path for random-sampling (never-ending) selectors.
                // Pull moves in small batches to avoid holding an immutable borrow
                // on the score director across the mutable evaluation calls.
                // The selector provides its own randomization; no shuffle needed.
                'streaming: loop {
                    let batch_start = self.arena.len();
                    {
                        let iter = self.move_selector.iter_moves(step_scope.score_director());
                        for mov in iter.take(64) {
                            self.arena.push(mov);
                        }
                    }
                    let batch_end = self.arena.len();
                    if batch_end == batch_start {
                        break;
                    }

                    for i in batch_start..batch_end {
                        local_moves_evaluated += 1;

                        // Log progress every ~8k moves (avoids Instant::now() syscall per move)
                        if local_moves_evaluated & 0x1FFF == 0 {
                            let now = Instant::now();
                            if now.duration_since(last_progress_time).as_secs() >= 1 {
                                let moves_delta = local_moves_evaluated - last_progress_moves;
                                let elapsed_secs =
                                    now.duration_since(last_progress_time).as_secs_f64();
                                let current_speed = (moves_delta as f64 / elapsed_secs) as u64;
                                debug!(
                                    event = "progress",
                                    steps = step_scope.step_index(),
                                    moves_evaluated = local_moves_evaluated,
                                    speed = current_speed,
                                    score = %last_step_score,
                                );
                                last_progress_time = now;
                                last_progress_moves = local_moves_evaluated;
                            }
                        }

                        let m = self.arena.get(i).unwrap();

                        if !m.is_doable(step_scope.score_director()) {
                            step_scope
                                .phase_scope_mut()
                                .solver_scope_mut()
                                .stats_mut()
                                .record_move(false);
                            continue;
                        }

                        let move_score = {
                            let mut recording =
                                RecordingScoreDirector::new(step_scope.score_director_mut());
                            m.do_move(&mut recording);
                            let score = recording.calculate_score();
                            recording.undo_changes();
                            score
                        };

                        step_scope
                            .phase_scope_mut()
                            .solver_scope_mut()
                            .stats_mut()
                            .record_score_calculation();

                        let accepted = self.acceptor.is_accepted(&last_step_score, &move_score);

                        step_scope
                            .phase_scope_mut()
                            .solver_scope_mut()
                            .stats_mut()
                            .record_move(accepted);

                        trace!(
                            event = "step",
                            step = step_scope.step_index(),
                            move_index = i,
                            score = %move_score,
                            accepted = accepted,
                        );

                        if accepted {
                            self.forager.add_move_index(i, move_score);
                        }

                        if self.forager.is_quit_early() {
                            break 'streaming;
                        }
                    }
                }
            } else {
                // Batch path: materialize full move space, shuffle, evaluate.
                self.arena
                    .extend(self.move_selector.iter_moves(step_scope.score_director()));

                // Shuffle arena in-place — O(n) Fisher-Yates, no allocation
                self.arena.shuffle(&mut rng);

                // Evaluate moves by index
                for i in 0..self.arena.len() {
                    local_moves_evaluated += 1;

                    // Log progress every ~8k moves (avoids Instant::now() syscall per move)
                    if local_moves_evaluated & 0x1FFF == 0 {
                        let now = Instant::now();
                        if now.duration_since(last_progress_time).as_secs() >= 1 {
                            let moves_delta = local_moves_evaluated - last_progress_moves;
                            let elapsed_secs = now.duration_since(last_progress_time).as_secs_f64();
                            let current_speed = (moves_delta as f64 / elapsed_secs) as u64;
                            debug!(
                                event = "progress",
                                steps = step_scope.step_index(),
                                moves_evaluated = local_moves_evaluated,
                                speed = current_speed,
                                score = %last_step_score,
                            );
                            last_progress_time = now;
                            last_progress_moves = local_moves_evaluated;
                        }
                    }

                    let m = self.arena.get(i).unwrap();

                    if !m.is_doable(step_scope.score_director()) {
                        // Record as evaluated but not accepted
                        step_scope
                            .phase_scope_mut()
                            .solver_scope_mut()
                            .stats_mut()
                            .record_move(false);
                        continue;
                    }

                    // Use RecordingScoreDirector for automatic undo
                    // This correctly handles state rollback for all moves including
                    // accepted-but-not-improving sidesteps (>= acceptance)
                    let move_score = {
                        let mut recording =
                            RecordingScoreDirector::new(step_scope.score_director_mut());

                        // Execute move
                        m.do_move(&mut recording);

                        // Calculate resulting score
                        let score = recording.calculate_score();

                        // Undo the move - state is fully restored regardless of acceptance
                        recording.undo_changes();

                        score
                    };

                    // Record score calculation (RecordingScoreDirector bypasses scope interceptor)
                    step_scope
                        .phase_scope_mut()
                        .solver_scope_mut()
                        .stats_mut()
                        .record_score_calculation();

                    // Check if accepted (>= allows sidesteps for plateau exploration)
                    let accepted = self.acceptor.is_accepted(&last_step_score, &move_score);

                    // Record move evaluation in solver stats
                    step_scope
                        .phase_scope_mut()
                        .solver_scope_mut()
                        .stats_mut()
                        .record_move(accepted);

                    trace!(
                        event = "step",
                        step = step_scope.step_index(),
                        move_index = i,
                        score = %move_score,
                        accepted = accepted,
                    );

                    // Add index to forager if accepted (not the move itself)
                    if accepted {
                        self.forager.add_move_index(i, move_score);
                    }

                    // Check if forager wants to quit early
                    if self.forager.is_quit_early() {
                        break;
                    }
                }
            }

            // Pick the best accepted move index
            if let Some((selected_index, selected_score)) = self.forager.pick_move_index() {
                // Execute the selected move permanently (by reference — no ownership needed).
                // The RecordingScoreDirector already undid the evaluation,
                // so this is a fresh application of the chosen move.
                self.arena
                    .get(selected_index)
                    .unwrap()
                    .do_move(step_scope.score_director_mut());
                step_scope.set_step_score(selected_score);

                // Update last step score
                last_step_score = selected_score;

                // Update best solution if improved
                step_scope.phase_scope_mut().update_best_solution();
            }
            // else: no accepted moves this step — that's fine, the acceptor
            // history still needs to advance so Late Acceptance / SA / etc.
            // can eventually escape the local optimum.

            // Always notify acceptor that step ended. For stateful acceptors
            // (Late Acceptance, Simulated Annealing, Great Deluge, SCHC),
            // the history must advance every step — even steps where no move
            // was accepted — otherwise the acceptor state stalls.
            self.acceptor.step_ended(&last_step_score);

            step_scope.complete();
        }

        // Notify acceptor of phase end
        self.acceptor.phase_ended();

        let duration = start_time.elapsed();
        let steps = phase_scope.step_count();
        let secs = duration.as_secs_f64();
        let speed = if secs > 0.0 {
            (local_moves_evaluated as f64 / secs) as u64
        } else {
            0
        };

        let stats = phase_scope.solver_scope().stats();
        let acceptance_rate = stats.acceptance_rate() * 100.0;
        let calc_speed = if secs > 0.0 {
            (stats.score_calculations as f64 / secs) as u64
        } else {
            0
        };

        let best_score_str = phase_scope
            .solver_scope()
            .best_score()
            .map(|s| format!("{}", s))
            .unwrap_or_else(|| "none".to_string());

        info!(
            event = "phase_end",
            phase = "Local Search",
            phase_index = phase_index,
            duration_ms = duration.as_millis() as u64,
            steps = steps,
            moves_speed = speed,
            calc_speed = calc_speed,
            acceptance_rate = format!("{:.1}%", acceptance_rate),
            score = best_score_str,
        );
    }

    fn phase_type_name(&self) -> &'static str {
        "LocalSearch"
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::heuristic::selector::ChangeMoveSelector;
    use crate::phase::localsearch::{AcceptedCountForager, HillClimbingAcceptor};
    use crate::test_utils::{
        create_nqueens_director, get_queen_row, set_queen_row, NQueensSolution,
    };
    use solverforge_core::score::SimpleScore;

    type NQueensMove = crate::heuristic::r#move::ChangeMove<NQueensSolution, i64>;

    fn create_move_selector(
        values: Vec<i64>,
    ) -> ChangeMoveSelector<
        NQueensSolution,
        i64,
        crate::heuristic::selector::FromSolutionEntitySelector,
        crate::heuristic::selector::StaticTypedValueSelector<NQueensSolution, i64>,
    > {
        ChangeMoveSelector::simple(get_queen_row, set_queen_row, 0, "row", values)
    }

    #[test]
    fn test_local_search_hill_climbing() {
        let director = create_nqueens_director(&[0, 0, 0, 0]);
        let mut solver_scope = SolverScope::new(director);
        solver_scope.start_solving();

        let initial_score = solver_scope.calculate_score();
        assert!(initial_score < SimpleScore::of(0));

        let values: Vec<i64> = (0..4).collect();
        let move_selector = create_move_selector(values);
        let acceptor = HillClimbingAcceptor::new();
        let forager: AcceptedCountForager<_> = AcceptedCountForager::new(1);
        let mut phase: LocalSearchPhase<_, NQueensMove, _, _, _> =
            LocalSearchPhase::new(move_selector, acceptor, forager, Some(100));

        phase.solve(&mut solver_scope);

        let final_score = solver_scope.best_score().copied().unwrap_or(initial_score);
        assert!(final_score >= initial_score);

        // Verify stats were recorded
        assert!(solver_scope.stats().moves_evaluated > 0);
    }

    #[test]
    fn test_local_search_reaches_optimal() {
        let director = create_nqueens_director(&[0, 2, 1, 3]);
        let mut solver_scope = SolverScope::new(director);
        solver_scope.start_solving();

        let initial_score = solver_scope.calculate_score();

        let values: Vec<i64> = (0..4).collect();
        let move_selector = create_move_selector(values);
        let acceptor = HillClimbingAcceptor::new();
        let forager: AcceptedCountForager<_> = AcceptedCountForager::new(1);
        let mut phase: LocalSearchPhase<_, NQueensMove, _, _, _> =
            LocalSearchPhase::new(move_selector, acceptor, forager, Some(50));

        phase.solve(&mut solver_scope);

        let final_score = solver_scope.best_score().copied().unwrap_or(initial_score);
        assert!(final_score >= initial_score);
    }

    #[test]
    fn test_local_search_step_limit() {
        let director = create_nqueens_director(&[0, 0, 0, 0]);
        let mut solver_scope = SolverScope::new(director);
        solver_scope.start_solving();

        let values: Vec<i64> = (0..4).collect();
        let move_selector = create_move_selector(values);
        let acceptor = HillClimbingAcceptor::new();
        let forager: AcceptedCountForager<_> = AcceptedCountForager::new(1);
        let mut phase: LocalSearchPhase<_, NQueensMove, _, _, _> =
            LocalSearchPhase::new(move_selector, acceptor, forager, Some(3));

        phase.solve(&mut solver_scope);

        // Steps should be limited
        assert!(solver_scope.stats().step_count <= 3);
    }
}