optirs_core/online_learning/

mod.rs

// Online learning and lifelong optimization
//
// This module provides optimization strategies for continuous learning scenarios,
// including online learning, continual learning, and lifelong optimization systems.

use crate::error::{OptimError, Result};
use crate::utils::{scalar_or, total_order, try_f64, try_scalar};
use scirs2_core::ndarray::{Array, Dimension, ScalarOperand};
use scirs2_core::numeric::Float;
use scirs2_core::random::thread_rng;
use std::collections::{HashMap, VecDeque};
use std::fmt::Debug;

mod transfer;

pub use transfer::{
    cosine_similarity, TaskStatistics, TransferOutcome, DEFAULT_TASK_EMBEDDING_DIM,
    DEFAULT_TRANSFER_THRESHOLD, MAX_TASK_EMBEDDING_DIM,
};

/// Online learning strategy
#[derive(Debug, Clone)]
pub enum OnlineLearningStrategy {
    /// Stochastic Gradient Descent with adaptive learning rate
    AdaptiveSGD {
        /// Initial learning rate
        initial_lr: f64,
        /// Learning rate adaptation method
        adaptation_method: LearningRateAdaptation,
    },
    /// Online Newton's method (second-order)
    OnlineNewton {
        /// Damping parameter for stability
        damping: f64,
        /// Window size for Hessian estimation
        hessian_window: usize,
    },
    /// Follow The Regularized Leader (FTRL)
    FTRL {
        /// L1 regularization strength
        l1_regularization: f64,
        /// L2 regularization strength
        l2_regularization: f64,
        /// Learning rate power
        learning_rate_power: f64,
    },
    /// Online Mirror Descent
    MirrorDescent {
        /// Mirror function type
        mirror_function: MirrorFunction,
        /// Regularization strength
        regularization: f64,
    },
    /// Adaptive Multi-Task Learning
    AdaptiveMultiTask {
        /// Task similarity threshold
        similarity_threshold: f64,
        /// Task-specific learning rates
        task_lr_adaptation: bool,
    },
}

/// Learning rate adaptation methods for online learning
#[derive(Debug, Clone)]
pub enum LearningRateAdaptation {
    /// AdaGrad-style adaptation
    AdaGrad {
        /// Small constant for numerical stability
        epsilon: f64,
    },
    /// RMSprop-style adaptation
    RMSprop {
        /// Decay rate
        decay: f64,
        /// Small constant for numerical stability
        epsilon: f64,
    },
    /// Adam-style adaptation
    Adam {
        /// Exponential decay rate for first moment
        beta1: f64,
        /// Exponential decay rate for second moment
        beta2: f64,
        /// Small constant for numerical stability
        epsilon: f64,
    },
    /// Exponential decay
    ExponentialDecay {
        /// Decay rate
        decay_rate: f64,
    },
    /// Inverse scaling
    InverseScaling {
        /// Scaling power
        power: f64,
    },
}

/// Mirror functions for mirror descent
#[derive(Debug, Clone)]
pub enum MirrorFunction {
    /// Euclidean (L2) regularization
    Euclidean,
    /// Entropy regularization (for probability simplex)
    Entropy,
    /// L1 regularization
    L1,
    /// Nuclear norm (for matrices)
    Nuclear,
}

/// Lifelong learning strategy
#[derive(Debug, Clone)]
pub enum LifelongStrategy {
    /// Elastic Weight Consolidation (EWC)
    ElasticWeightConsolidation {
        /// Importance weight for previous tasks
        importance_weight: f64,
        /// Fisher information estimation samples
        fisher_samples: usize,
    },
    /// Progressive Neural Networks
    ProgressiveNetworks {
        /// Lateral connection strength
        lateral_strength: f64,
        /// Column growth strategy
        growth_strategy: ColumnGrowthStrategy,
    },
    /// Memory-Augmented Networks
    MemoryAugmented {
        /// Memory size
        memory_size: usize,
        /// Memory update strategy
        update_strategy: MemoryUpdateStrategy,
    },
    /// Meta-Learning Based Continual Learning
    MetaLearning {
        /// Meta-learning rate
        meta_lr: f64,
        /// Inner loop steps
        inner_steps: usize,
        /// Task embedding size
        task_embedding_size: usize,
    },
    /// Gradient Episodic Memory (GEM)
    GradientEpisodicMemory {
        /// Memory buffer size per task
        memory_per_task: usize,
        /// Constraint violation tolerance
        violation_tolerance: f64,
    },
}

/// Column growth strategies for progressive networks
#[derive(Debug, Clone)]
pub enum ColumnGrowthStrategy {
    /// Add new column for each task
    PerTask,
    /// Add new column when performance drops
    PerformanceBased {
        /// Performance threshold
        threshold: f64,
    },
    /// Add new column after fixed intervals
    FixedInterval {
        /// Fixed interval
        interval: usize,
    },
}

/// Memory update strategies
#[derive(Debug, Clone)]
pub enum MemoryUpdateStrategy {
    /// First In First Out
    FIFO,
    /// Random replacement
    Random,
    /// Importance-based replacement
    ImportanceBased,
    /// Gradient diversity
    GradientDiversity,
}

/// Online optimizer that adapts to streaming data
#[derive(Debug)]
pub struct OnlineOptimizer<A: Float, D: Dimension> {
    /// Online learning strategy
    strategy: OnlineLearningStrategy,
    /// Current parameters
    parameters: Array<A, D>,
    /// Accumulated gradients for adaptation
    gradient_accumulator: Array<A, D>,
    /// Second moment accumulator (for Adam-like methods)
    second_moment_accumulator: Option<Array<A, D>>,
    /// Current learning rate
    current_lr: A,
    /// Step counter
    step_count: usize,
    /// Performance history
    performance_history: VecDeque<A>,
    /// Regret bounds tracking
    regret_bound: A,
    /// Bounded history of the (possibly adapted) learning rate, used to
    /// compute a real `lr_stability` metric (F81).
    lr_history: VecDeque<A>,
}

/// Lifelong optimizer that learns continuously across tasks
#[derive(Debug)]
pub struct LifelongOptimizer<A: Float, D: Dimension> {
    /// Lifelong learning strategy
    strategy: LifelongStrategy,
    /// Task-specific optimizers
    task_optimizers: HashMap<String, OnlineOptimizer<A, D>>,
    /// Shared knowledge across tasks
    shared_knowledge: SharedKnowledge<A, D>,
    /// Task sequence and relationships
    task_graph: TaskGraph,
    /// Memory buffer for important examples
    memory_buffer: MemoryBuffer<A, D>,
    /// Current active task
    current_task: Option<String>,
    /// Performance tracking across tasks
    task_performance: HashMap<String, Vec<A>>,
    /// Last loss each task recorded while it was still the active task: the
    /// reference a later re-evaluation is compared against to measure
    /// catastrophic forgetting.
    task_reference_loss: HashMap<String, A>,
    /// Cosine similarity a new task must reach before it is warm-started from
    /// an existing one, and the linkage at which tasks are clustered.
    transfer_threshold: f64,
}

/// Shared knowledge representation for lifelong learning
#[derive(Debug)]
pub struct SharedKnowledge<A: Float, D: Dimension> {
    /// Diagonal empirical Fisher information `F_i = E[g_i^2]`, accumulated
    /// across the gradients observed for the current task (EWC).
    fisher_information: Option<Array<A, D>>,
    /// Number of gradients folded into `fisher_information` so far, capped at
    /// the strategy's `fisher_samples` so the estimate keeps tracking as an
    /// exponential moving average afterwards.
    fisher_sample_count: usize,
    /// Consolidated parameters `θ*` of the previously-learned tasks: the anchor
    /// the EWC penalty pulls the active task back toward.
    important_parameters: Option<Array<A, D>>,
    /// Meta-parameters shared across tasks, maintained by the first-order
    /// Reptile update in [`LifelongOptimizer::apply_meta_learning`].
    meta_parameters: Option<Array<A, D>>,
    /// Running gradient statistics per task, the raw material every task
    /// embedding is derived from.
    task_statistics: HashMap<String, TaskStatistics>,
    /// Fixed-width task embeddings derived from `task_statistics`.
    task_embeddings: HashMap<String, Vec<f64>>,
    /// Warm-start strength actually applied, keyed by `(source, target)`.
    transfer_weights: HashMap<(String, String), f64>,
}

/// Task relationship graph
#[derive(Debug)]
pub struct TaskGraph {
    /// Task relationships (similarity scores)
    task_similarities: HashMap<(String, String), f64>,
    /// Tasks each task was warm-started from.
    task_dependencies: HashMap<String, Vec<String>>,
    /// Agglomerative clustering of `task_similarities`.
    task_clusters: Vec<Vec<String>>,
}

/// Memory buffer for important examples
#[derive(Debug)]
pub struct MemoryBuffer<A: Float, D: Dimension> {
    /// Stored examples
    examples: VecDeque<MemoryExample<A, D>>,
    /// Maximum buffer size
    max_size: usize,
    /// Update strategy
    update_strategy: MemoryUpdateStrategy,
    /// Importance scores
    importance_scores: VecDeque<A>,
}

/// Single memory example
#[derive(Debug, Clone)]
pub struct MemoryExample<A: Float, D: Dimension> {
    /// Input data
    pub input: Array<A, D>,
    /// Target output
    pub target: Array<A, D>,
    /// Task identifier
    pub task_id: String,
    /// Importance score
    pub importance: A,
    /// Gradient information
    pub gradient: Option<Array<A, D>>,
}

/// Online learning performance metrics
#[derive(Debug, Clone)]
pub struct OnlinePerformanceMetrics<A: Float> {
    /// Cumulative regret
    pub cumulative_regret: A,
    /// Average loss over window
    pub average_loss: A,
    /// Learning rate stability
    pub lr_stability: A,
    /// Adaptation speed
    pub adaptation_speed: A,
    /// Memory efficiency
    pub memory_efficiency: A,
}

impl<A: Float + ScalarOperand + Debug + std::iter::Sum, D: Dimension + Send + Sync>
    OnlineOptimizer<A, D>
{
    /// Create a new online optimizer
    pub fn new(strategy: OnlineLearningStrategy, initial_parameters: Array<A, D>) -> Self {
        let paramshape = initial_parameters.raw_dim();
        let gradient_accumulator = Array::zeros(paramshape.clone());
        let second_moment_accumulator = match &strategy {
            OnlineLearningStrategy::AdaptiveSGD {
                adaptation_method: LearningRateAdaptation::Adam { .. },
                ..
            } => Some(Array::zeros(paramshape)),
            _ => None,
        };

        let current_lr = match &strategy {
            OnlineLearningStrategy::AdaptiveSGD { initial_lr, .. } => {
                scalar_or(*initial_lr, A::zero())
            }
            OnlineLearningStrategy::OnlineNewton { .. } => scalar_or(0.01, A::zero()),
            OnlineLearningStrategy::FTRL { .. } => scalar_or(0.1, A::zero()),
            OnlineLearningStrategy::MirrorDescent { .. } => scalar_or(0.01, A::zero()),
            OnlineLearningStrategy::AdaptiveMultiTask { .. } => scalar_or(0.001, A::zero()),
        };

        Self {
            strategy,
            parameters: initial_parameters,
            gradient_accumulator,
            second_moment_accumulator,
            current_lr,
            step_count: 0,
            performance_history: VecDeque::new(),
            regret_bound: A::zero(),
            lr_history: VecDeque::new(),
        }
    }

    /// Perform online update with new gradient
    pub fn online_update(&mut self, gradient: &Array<A, D>, loss: A) -> Result<()> {
        self.step_count += 1;
        self.performance_history.push_back(loss);

        // Keep performance history bounded
        if self.performance_history.len() > 1000 {
            self.performance_history.pop_front();
        }

        match self.strategy.clone() {
            OnlineLearningStrategy::AdaptiveSGD {
                adaptation_method, ..
            } => {
                self.adaptive_sgd_update(gradient, &adaptation_method)?;
            }
            OnlineLearningStrategy::OnlineNewton { damping, .. } => {
                self.online_newton_update(gradient, damping)?;
            }
            OnlineLearningStrategy::FTRL {
                l1_regularization,
                l2_regularization,
                learning_rate_power,
            } => {
                self.ftrl_update(
                    gradient,
                    l1_regularization,
                    l2_regularization,
                    learning_rate_power,
                )?;
            }
            OnlineLearningStrategy::MirrorDescent {
                mirror_function,
                regularization,
            } => {
                self.mirror_descent_update(gradient, &mirror_function, regularization)?;
            }
            OnlineLearningStrategy::AdaptiveMultiTask { .. } => {
                self.adaptive_multitask_update(gradient)?;
            }
        }

        // Record the (possibly adapted) learning rate for the stability
        // metric (F81).
        self.lr_history.push_back(self.current_lr);
        if self.lr_history.len() > 1000 {
            self.lr_history.pop_front();
        }

        // Update regret bound
        self.update_regret_bound(loss);

        Ok(())
    }

    /// Adaptive SGD update
    fn adaptive_sgd_update(
        &mut self,
        gradient: &Array<A, D>,
        adaptation: &LearningRateAdaptation,
    ) -> Result<()> {
        match adaptation {
            LearningRateAdaptation::AdaGrad { epsilon } => {
                // Accumulate squared gradients
                self.gradient_accumulator = &self.gradient_accumulator + &gradient.mapv(|g| g * g);

                // Compute adaptive learning rate
                // Hoisted out of the closure: the conversion is loop-invariant and
                // `?` cannot cross a closure boundary.
                let eps = try_scalar::<A, _>(*epsilon)?;
                let adaptive_lr = self.gradient_accumulator.mapv(|acc| eps + A::sqrt(acc));

                // Update parameters
                self.parameters = &self.parameters - &(gradient / &adaptive_lr * self.current_lr);
            }
            LearningRateAdaptation::RMSprop { decay, epsilon } => {
                let decay_factor = try_scalar::<A, _>(*decay)?;
                let one_minus_decay = A::one() - decay_factor;

                // Update moving average of squared gradients
                self.gradient_accumulator = &self.gradient_accumulator * decay_factor
                    + &gradient.mapv(|g| g * g * one_minus_decay);

                // Compute adaptive learning rate
                let eps = try_scalar::<A, _>(*epsilon)?;
                let adaptive_lr = self.gradient_accumulator.mapv(|acc| A::sqrt(acc + eps));

                // Update parameters
                self.parameters = &self.parameters - &(gradient / &adaptive_lr * self.current_lr);
            }
            LearningRateAdaptation::Adam {
                beta1,
                beta2,
                epsilon,
            } => {
                let beta1_val = try_scalar::<A, _>(*beta1)?;
                let beta2_val = try_scalar::<A, _>(*beta2)?;
                let one_minus_beta1 = A::one() - beta1_val;
                let one_minus_beta2 = A::one() - beta2_val;

                // Update first moment (gradient accumulator)
                self.gradient_accumulator =
                    &self.gradient_accumulator * beta1_val + gradient * one_minus_beta1;

                // Update second moment
                if let Some(ref mut second_moment) = self.second_moment_accumulator {
                    *second_moment =
                        &*second_moment * beta2_val + &gradient.mapv(|g| g * g * one_minus_beta2);

                    // Bias correction
                    let step_count_float = try_scalar::<A, _>(self.step_count)?;
                    let bias_correction1 = A::one() - A::powf(beta1_val, step_count_float);
                    let bias_correction2 = A::one() - A::powf(beta2_val, step_count_float);

                    let corrected_first = &self.gradient_accumulator / bias_correction1;
                    let corrected_second = &*second_moment / bias_correction2;

                    // Update parameters
                    let eps = try_scalar::<A, _>(*epsilon)?;
                    let adaptive_lr = corrected_second.mapv(|v| A::sqrt(v) + eps);
                    self.parameters =
                        &self.parameters - &(corrected_first / adaptive_lr * self.current_lr);
                }
            }
            LearningRateAdaptation::ExponentialDecay { decay_rate } => {
                // Simple exponential decay
                self.current_lr = self.current_lr * try_scalar::<A, _>(*decay_rate)?;
                self.parameters = &self.parameters - gradient * self.current_lr;
            }
            LearningRateAdaptation::InverseScaling { power } => {
                // Inverse scaling: lr = initial_lr / (step^power)
                let step_power = A::powf(
                    try_scalar::<A, _>(self.step_count)?,
                    try_scalar::<A, _>(*power)?,
                );
                let decayed_lr = self.current_lr / step_power;
                self.parameters = &self.parameters - gradient * decayed_lr;
            }
        }

        Ok(())
    }

    /// Online Newton's method update
    fn online_newton_update(&mut self, gradient: &Array<A, D>, damping: f64) -> Result<()> {
        // Simplified online Newton update with damping
        let damping_val = try_scalar::<A, _>(damping)?;

        // Approximate Hessian diagonal with gradient squares (simplified)
        let hessian_approx = gradient.mapv(|g| g * g + damping_val);

        // Newton step
        let newton_step = gradient / hessian_approx;
        self.parameters = &self.parameters - &newton_step * self.current_lr;

        Ok(())
    }

    /// FTRL update
    fn ftrl_update(
        &mut self,
        gradient: &Array<A, D>,
        l1_reg: f64,
        l2_reg: f64,
        lr_power: f64,
    ) -> Result<()> {
        // Accumulate gradients
        self.gradient_accumulator = &self.gradient_accumulator + gradient;

        // FTRL update rule (simplified)
        let step_factor = A::powf(
            try_scalar::<A, _>(self.step_count)?,
            try_scalar::<A, _>(lr_power)?,
        );
        let learning_rate = self.current_lr / step_factor;

        // Apply L1 and L2 regularization
        let l1_weight = try_scalar::<A, _>(l1_reg)?;
        let l2_weight = try_scalar::<A, _>(l2_reg)?;

        self.parameters = self.gradient_accumulator.mapv(|g| {
            let abs_g = A::abs(g);
            if abs_g <= l1_weight {
                A::zero()
            } else {
                let sign = if g > A::zero() { A::one() } else { -A::one() };
                -sign * (abs_g - l1_weight) / (l2_weight + A::sqrt(abs_g))
            }
        }) * learning_rate;

        Ok(())
    }

    /// Mirror descent update
    fn mirror_descent_update(
        &mut self,
        gradient: &Array<A, D>,
        mirror_fn: &MirrorFunction,
        regularization: f64,
    ) -> Result<()> {
        match mirror_fn {
            MirrorFunction::Euclidean => {
                // Standard gradient descent
                self.parameters = &self.parameters - gradient * self.current_lr;
            }
            MirrorFunction::Entropy => {
                // Entropy regularized update (for probability simplex)
                let reg_val = try_scalar::<A, _>(regularization)?;
                let updated = self
                    .parameters
                    .mapv(|p| A::exp(A::ln(p) - self.current_lr * reg_val));
                let sum = updated.sum();
                self.parameters = updated / sum; // Normalize to probability simplex
            }
            MirrorFunction::L1 => {
                // L1 regularized update with soft thresholding
                let threshold = self.current_lr * try_scalar::<A, _>(regularization)?;
                self.parameters = (&self.parameters - gradient * self.current_lr).mapv(|p| {
                    if A::abs(p) <= threshold {
                        A::zero()
                    } else {
                        p - A::signum(p) * threshold
                    }
                });
            }
            MirrorFunction::Nuclear => {
                // Simplified nuclear norm update (requires matrix structure)
                self.parameters = &self.parameters - gradient * self.current_lr;
            }
        }

        Ok(())
    }

    /// Adaptive multi-task update
    fn adaptive_multitask_update(&mut self, gradient: &Array<A, D>) -> Result<()> {
        // Simplified multi-task update
        self.parameters = &self.parameters - gradient * self.current_lr;
        Ok(())
    }

    /// Update regret bound estimation
    fn update_regret_bound(&mut self, loss: A) {
        if let Some(&best_loss) = self
            .performance_history
            .iter()
            .min_by(|a, b| total_order(*a, *b))
        {
            let regret = loss - best_loss;
            self.regret_bound = self.regret_bound + regret.max(A::zero());
        }
    }

    /// Get current parameters
    pub fn parameters(&self) -> &Array<A, D> {
        &self.parameters
    }

    /// Get performance metrics
    pub fn get_performance_metrics(&self) -> OnlinePerformanceMetrics<A> {
        let average_loss = if self.performance_history.is_empty() {
            A::zero()
        } else {
            self.performance_history.iter().copied().sum::<A>()
                / A::from(self.performance_history.len()).unwrap_or_else(A::one)
        };

        // Learning-rate stability (F81): high when the recent learning rate
        // has low variance, mapped into `(0, 1]` via `1 / (1 + std(lr))`.
        // Previously a hardcoded `1.0`.
        let lr_stability = {
            let n = self.lr_history.len();
            if n < 2 {
                A::one()
            } else {
                let count = A::from(n).unwrap_or_else(A::one);
                let mean = self.lr_history.iter().copied().sum::<A>() / count;
                let variance = self
                    .lr_history
                    .iter()
                    .map(|&lr| (lr - mean) * (lr - mean))
                    .sum::<A>()
                    / count;
                A::one() / (A::one() + variance.sqrt())
            }
        };

        // Adaptation speed (F81): mean per-step loss improvement over a
        // recent window (positive = loss decreasing), clamped at 0.
        // Previously the raw `step_count`, which is a count, not a speed.
        let adaptation_speed = {
            let history_len = self.performance_history.len();
            let window = 20usize.min(history_len);
            if window < 2 {
                A::zero()
            } else {
                let first = self.performance_history[history_len - window];
                let last = self.performance_history[history_len - 1];
                let steps = A::from(window - 1).unwrap_or_else(A::one);
                ((first - last) / steps).max(A::zero())
            }
        };

        // Memory efficiency (F81): parameter storage as a fraction of the
        // optimizer's total array storage (parameters + accumulators).
        // Higher means less auxiliary-state overhead. Previously a hardcoded
        // `0.8`.
        let memory_efficiency = {
            let param_len = self.parameters.len();
            let mut total = param_len + self.gradient_accumulator.len();
            if let Some(second_moment) = &self.second_moment_accumulator {
                total += second_moment.len();
            }
            if total == 0 {
                A::zero()
            } else {
                A::from(param_len).unwrap_or_else(A::one) / A::from(total).unwrap_or_else(A::one)
            }
        };

        OnlinePerformanceMetrics {
            cumulative_regret: self.regret_bound,
            average_loss,
            lr_stability,
            adaptation_speed,
            memory_efficiency,
        }
    }
}

/// Euclidean dot product of two equally-shaped arrays. Used by the
/// Gradient Episodic Memory projection (F81).
fn array_dot<A: Float + std::iter::Sum, D: Dimension>(a: &Array<A, D>, b: &Array<A, D>) -> A {
    a.iter().zip(b.iter()).map(|(&x, &y)| x * y).sum()
}

impl<A: Float + ScalarOperand + Debug + std::iter::Sum, D: Dimension + Send + Sync>
    LifelongOptimizer<A, D>
{
    /// Create a new lifelong optimizer
    pub fn new(strategy: LifelongStrategy) -> Self {
        Self {
            strategy,
            task_optimizers: HashMap::new(),
            shared_knowledge: SharedKnowledge {
                fisher_information: None,
                fisher_sample_count: 0,
                important_parameters: None,
                meta_parameters: None,
                task_statistics: HashMap::new(),
                task_embeddings: HashMap::new(),
                transfer_weights: HashMap::new(),
            },
            task_graph: TaskGraph {
                task_similarities: HashMap::new(),
                task_dependencies: HashMap::new(),
                task_clusters: Vec::new(),
            },
            memory_buffer: MemoryBuffer {
                examples: VecDeque::new(),
                max_size: 1000,
                update_strategy: MemoryUpdateStrategy::FIFO,
                importance_scores: VecDeque::new(),
            },
            current_task: None,
            task_performance: HashMap::new(),
            task_reference_loss: HashMap::new(),
            transfer_threshold: DEFAULT_TRANSFER_THRESHOLD,
        }
    }

    /// Start learning a new task
    ///
    /// Before switching, the outgoing task is *consolidated*: its final
    /// parameters become the shared anchor `θ*` that the Elastic Weight
    /// Consolidation penalty pulls subsequent tasks back toward, and the
    /// Fisher-information sample counter is reset so the next task's estimate
    /// starts from its own gradients rather than inheriting the old average.
    /// The outgoing task's last loss is also pinned as the reference a later
    /// re-evaluation is compared against ([`Self::record_task_performance`]).
    pub fn start_task(&mut self, task_id: String, initial_parameters: Array<A, D>) -> Result<()> {
        if let Some(previous) = self.current_task.clone() {
            if let Some(optimizer) = self.task_optimizers.get(&previous) {
                self.shared_knowledge.important_parameters = Some(optimizer.parameters().clone());
                self.shared_knowledge.fisher_sample_count = 0;
            }
            if let Some(&last_loss) = self
                .task_performance
                .get(&previous)
                .and_then(|history| history.last())
            {
                self.task_reference_loss.insert(previous, last_loss);
            }
        }

        self.current_task = Some(task_id.clone());

        // Create task-specific optimizer
        let online_strategy = OnlineLearningStrategy::AdaptiveSGD {
            initial_lr: 0.001,
            adaptation_method: LearningRateAdaptation::Adam {
                beta1: 0.9,
                beta2: 0.999,
                epsilon: 1e-8,
            },
        };

        let task_optimizer = OnlineOptimizer::new(online_strategy, initial_parameters);
        self.task_optimizers.insert(task_id.clone(), task_optimizer);

        // Initialize task performance tracking
        self.task_performance.insert(task_id, Vec::new());

        Ok(())
    }

    /// Update current task with new data
    pub fn update_current_task(&mut self, gradient: &Array<A, D>, loss: A) -> Result<()> {
        let task_id = self
            .current_task
            .as_ref()
            .ok_or_else(|| OptimError::InvalidConfig("No current task set".to_string()))?
            .clone();

        // For Gradient Episodic Memory, project the incoming gradient onto
        // the halfspace that does not increase loss on stored past-task
        // gradients BEFORE it is applied (F81). Other strategies use the raw
        // gradient. Projecting here (rather than in a no-op after the step)
        // is what makes the stored replay gradients actually shape learning.
        let is_gem = matches!(
            self.strategy,
            LifelongStrategy::GradientEpisodicMemory { .. }
        );
        let ewc_weight = match self.strategy {
            LifelongStrategy::ElasticWeightConsolidation {
                importance_weight, ..
            } => Some(importance_weight),
            _ => None,
        };
        let effective_gradient = if is_gem {
            self.project_gradient_gem(gradient)
        } else if let Some(importance_weight) = ewc_weight {
            // EWC: the total gradient is the task gradient plus the quadratic
            // penalty gradient `λ * F ⊙ (θ - θ*)`. Adding it *before* the step
            // (rather than in a no-op afterwards, as this used to) is what
            // makes `importance_weight`, `fisher_information` and
            // `important_parameters` actually shape learning.
            match self.ewc_penalty_gradient(importance_weight) {
                Some(penalty) => gradient + &penalty,
                None => gradient.clone(),
            }
        } else {
            gradient.clone()
        };

        // Update task-specific optimizer with the effective gradient
        if let Some(optimizer) = self.task_optimizers.get_mut(&task_id) {
            optimizer.online_update(&effective_gradient, loss)?;
        }

        // Track performance
        if let Some(performance) = self.task_performance.get_mut(&task_id) {
            performance.push(loss);
        }

        // Fold the task's *own* gradient (not the EWC/GEM-adjusted one, which
        // carries the regularizer rather than the task) into the running
        // statistics that drive cross-task transfer.
        self.record_task_observation(gradient)?;

        // Apply lifelong learning strategy
        match &self.strategy {
            LifelongStrategy::ElasticWeightConsolidation {
                importance_weight, ..
            } => {
                self.apply_ewc_regularization(gradient, *importance_weight)?;
            }
            LifelongStrategy::ProgressiveNetworks { .. } => {
                self.apply_progressive_networks(gradient)?;
            }
            LifelongStrategy::MemoryAugmented { .. } => {
                self.update_memory_buffer(gradient, loss)?;
            }
            LifelongStrategy::MetaLearning { .. } => {
                self.apply_meta_learning(gradient)?;
            }
            LifelongStrategy::GradientEpisodicMemory { .. } => {
                // The projection was already applied above; record this
                // step's *raw* gradient as an episodic-memory constraint for
                // future updates to be projected against.
                self.update_memory_buffer(gradient, loss)?;
            }
        }

        Ok(())
    }

    /// Fisher-information penalty gradient for Elastic Weight Consolidation.
    ///
    /// Returns `λ * F ⊙ (θ - θ*)`, the gradient of the EWC quadratic penalty
    /// `λ/2 * Σ F_i (θ_i - θ*_i)^2` (Kirkpatrick et al., "Overcoming
    /// catastrophic forgetting in neural networks", PNAS 2017), where `F` is
    /// the diagonal empirical Fisher accumulated for the current task and `θ*`
    /// is the consolidated parameter vector of the previously-learned tasks.
    ///
    /// Returns `None` — meaning "no penalty yet", not "failed" — while any of
    /// the three inputs is missing (before the first task has been
    /// consolidated there is nothing to forget) or when their shapes disagree,
    /// which can only happen if tasks of different parameter dimensionality
    /// are interleaved.
    fn ewc_penalty_gradient(&self, importance_weight: f64) -> Option<Array<A, D>> {
        let fisher = self.shared_knowledge.fisher_information.as_ref()?;
        let anchor = self.shared_knowledge.important_parameters.as_ref()?;
        let task_id = self.current_task.as_ref()?;
        let parameters = self.task_optimizers.get(task_id)?.parameters();

        if fisher.raw_dim() != anchor.raw_dim() || parameters.raw_dim() != anchor.raw_dim() {
            return None;
        }

        let lambda = scalar_or(importance_weight, A::zero());
        let mut penalty = Array::zeros(parameters.raw_dim());
        for (((out, &f), &p), &a) in penalty
            .iter_mut()
            .zip(fisher.iter())
            .zip(parameters.iter())
            .zip(anchor.iter())
        {
            *out = lambda * f * (p - a);
        }
        Some(penalty)
    }

    /// Fold the observed gradient into the diagonal empirical Fisher estimate
    /// used by Elastic Weight Consolidation.
    ///
    /// The diagonal Fisher of a log-likelihood model is `E[g_i^2]`, so it is
    /// estimated here as a running mean of the squared gradients seen for the
    /// current task. The averaging weight is `1/k` with `k` capped at the
    /// strategy's `fisher_samples`: up to that many observations this is an
    /// exact running mean, after which it becomes an exponential moving
    /// average with the same weight, so the estimate keeps tracking a
    /// non-stationary task instead of freezing.
    ///
    /// This replaces a no-op that ignored both its arguments and left
    /// `fisher_information` permanently `None`.
    fn apply_ewc_regularization(
        &mut self,
        gradient: &Array<A, D>,
        _importance_weight: f64,
    ) -> Result<()> {
        let fisher_samples = match self.strategy {
            LifelongStrategy::ElasticWeightConsolidation { fisher_samples, .. } => {
                fisher_samples.max(1)
            }
            _ => return Ok(()),
        };

        let shape_matches = self
            .shared_knowledge
            .fisher_information
            .as_ref()
            .is_some_and(|fisher| fisher.raw_dim() == gradient.raw_dim());

        if shape_matches {
            let count = (self.shared_knowledge.fisher_sample_count + 1).min(fisher_samples);
            // `count >= 1`, so this weight is always in (0, 1].
            let weight = A::one() / scalar_or(count as f64, A::one());
            if let Some(fisher) = self.shared_knowledge.fisher_information.as_mut() {
                for (f, &g) in fisher.iter_mut().zip(gradient.iter()) {
                    *f = *f + (g * g - *f) * weight;
                }
            }
            self.shared_knowledge.fisher_sample_count = count;
        } else {
            // First gradient for this parameter shape: seed the estimate with
            // `g^2` rather than averaging against a zero/stale accumulator.
            let mut fisher = Array::zeros(gradient.raw_dim());
            for (f, &g) in fisher.iter_mut().zip(gradient.iter()) {
                *f = g * g;
            }
            self.shared_knowledge.fisher_information = Some(fisher);
            self.shared_knowledge.fisher_sample_count = 1;
        }

        Ok(())
    }

    /// Apply the Progressive Neural Networks strategy.
    ///
    /// Not implemented, and deliberately reported as such rather than silently
    /// doing nothing: Progressive Networks (Rusu et al., arXiv:1606.04671)
    /// works by *growing a new network column per task and adding lateral
    /// connections from the frozen previous columns*. Both operations act on a
    /// model's layer graph, which an optimizer-only crate does not represent —
    /// there is no per-column parameter partition here to freeze, and
    /// `lateral_strength` has nothing to scale. Previously this returned
    /// `Ok(())` and ignored its argument, so selecting the strategy quietly
    /// produced plain online SGD while reporting success.
    fn apply_progressive_networks(&mut self, _gradient: &Array<A, D>) -> Result<()> {
        Err(OptimError::UnsupportedOperation(
            "LifelongStrategy::ProgressiveNetworks requires per-column model \
             parameter partitioning and lateral connections, which optirs-core \
             does not model; use ElasticWeightConsolidation, MemoryAugmented, \
             MetaLearning or GradientEpisodicMemory instead"
                .to_string(),
        ))
    }

    /// Update memory buffer with important examples
    fn update_memory_buffer(&mut self, gradient: &Array<A, D>, loss: A) -> Result<()> {
        if let Some(task_id) = &self.current_task {
            // `input`/`target` are not available at this call site (only the
            // gradient and loss are passed in), so they are recorded as
            // empty/zero and intentionally left unused. The `gradient` field
            // below is the real replay signal consumed by
            // `project_gradient_gem` (F81).
            let example = MemoryExample {
                input: Array::zeros(gradient.raw_dim()),
                target: Array::zeros(gradient.raw_dim()),
                task_id: task_id.clone(),
                importance: loss,
                gradient: Some(gradient.clone()),
            };

            // Add to buffer
            if self.memory_buffer.examples.len() >= self.memory_buffer.max_size {
                match self.memory_buffer.update_strategy {
                    MemoryUpdateStrategy::FIFO => {
                        self.memory_buffer.examples.pop_front();
                        self.memory_buffer.importance_scores.pop_front();
                    }
                    MemoryUpdateStrategy::Random => {
                        let idx = thread_rng().gen_range(0..self.memory_buffer.examples.len());
                        self.memory_buffer.examples.remove(idx);
                        self.memory_buffer.importance_scores.remove(idx);
                    }
                    MemoryUpdateStrategy::ImportanceBased => {
                        // Remove least important example
                        if let Some(min_idx) = self
                            .memory_buffer
                            .importance_scores
                            .iter()
                            .enumerate()
                            .min_by(|a, b| total_order(a.1, b.1))
                            .map(|(idx, _)| idx)
                        {
                            self.memory_buffer.examples.remove(min_idx);
                            self.memory_buffer.importance_scores.remove(min_idx);
                        }
                    }
                    MemoryUpdateStrategy::GradientDiversity => {
                        // Remove most similar gradient (simplified)
                        self.memory_buffer.examples.pop_front();
                        self.memory_buffer.importance_scores.pop_front();
                    }
                }
            }

            self.memory_buffer.examples.push_back(example);
            self.memory_buffer.importance_scores.push_back(loss);
        }

        Ok(())
    }

    /// Apply the meta-learning strategy: a first-order Reptile update of the
    /// shared meta-parameters.
    ///
    /// Reptile (Nichol, Achiam & Schulman, "On First-Order Meta-Learning
    /// Algorithms", arXiv:1803.02999) needs no second derivatives: after the
    /// inner-loop task update it simply moves the shared initialisation toward
    /// the adapted task parameters,
    /// `φ ← φ + meta_lr * (θ_task − φ)`,
    /// which is exactly what is available at this call site. The meta-parameters
    /// are seeded from the first task's parameters, so `φ` starts on the
    /// parameter manifold instead of at an arbitrary origin.
    ///
    /// `gradient` is used only for its shape (the parameter update itself has
    /// already been applied by the task optimizer, which is what makes this the
    /// *first-order* variant). This replaces a no-op that left
    /// `meta_parameters` permanently `None`.
    fn apply_meta_learning(&mut self, gradient: &Array<A, D>) -> Result<()> {
        let meta_lr = match self.strategy {
            LifelongStrategy::MetaLearning { meta_lr, .. } => meta_lr,
            _ => return Ok(()),
        };

        let Some(task_id) = self.current_task.clone() else {
            return Ok(());
        };
        let Some(optimizer) = self.task_optimizers.get(&task_id) else {
            return Ok(());
        };
        let task_parameters = optimizer.parameters();
        if task_parameters.raw_dim() != gradient.raw_dim() {
            return Err(OptimError::DimensionMismatch(format!(
                "meta-learning update: task parameters have shape {:?} but the \
                 gradient has shape {:?}",
                task_parameters.raw_dim().slice(),
                gradient.raw_dim().slice()
            )));
        }

        let shape_matches = self
            .shared_knowledge
            .meta_parameters
            .as_ref()
            .is_some_and(|meta| meta.raw_dim() == task_parameters.raw_dim());

        if shape_matches {
            let step = scalar_or(meta_lr, A::zero());
            let adapted = task_parameters.clone();
            if let Some(meta) = self.shared_knowledge.meta_parameters.as_mut() {
                for (phi, &theta) in meta.iter_mut().zip(adapted.iter()) {
                    *phi = *phi + (theta - *phi) * step;
                }
            }
        } else {
            self.shared_knowledge.meta_parameters = Some(task_parameters.clone());
        }

        Ok(())
    }

    /// Shared meta-parameters maintained by the Reptile update, if the
    /// `MetaLearning` strategy is active and at least one task has been seen.
    pub fn meta_parameters(&self) -> Option<&Array<A, D>> {
        self.shared_knowledge.meta_parameters.as_ref()
    }

    /// Diagonal empirical Fisher information accumulated for the current task,
    /// if the `ElasticWeightConsolidation` strategy is active and at least one
    /// gradient has been observed.
    pub fn fisher_information(&self) -> Option<&Array<A, D>> {
        self.shared_knowledge.fisher_information.as_ref()
    }

    /// Consolidated parameters `θ*` of the previously-learned tasks, set when
    /// [`Self::start_task`] switches away from a task.
    pub fn consolidated_parameters(&self) -> Option<&Array<A, D>> {
        self.shared_knowledge.important_parameters.as_ref()
    }

    /// Project `gradient` onto the halfspace that does not increase loss on
    /// any gradient stored in the episodic memory buffer (Gradient Episodic
    /// Memory, F81).
    ///
    /// For every stored memory gradient `g_mem` that conflicts with the
    /// incoming gradient (`dot(gradient, g_mem) < 0`), the violating
    /// component is removed:
    /// `g <- g - (dot(g, g_mem) / dot(g_mem, g_mem)) * g_mem`.
    /// Non-conflicting memories and shape-mismatched entries are skipped.
    ///
    /// This replaces the previous no-op and makes the replay gradients
    /// stored in the memory buffer actually influence learning instead of
    /// being dead data.
    pub fn project_gradient_gem(&self, gradient: &Array<A, D>) -> Array<A, D> {
        let mut projected = gradient.clone();
        for example in &self.memory_buffer.examples {
            if let Some(memory_gradient) = example.gradient.as_ref() {
                if memory_gradient.raw_dim() != projected.raw_dim() {
                    continue;
                }
                let dot_gm = array_dot(&projected, memory_gradient);
                if dot_gm < A::zero() {
                    let denom = array_dot(memory_gradient, memory_gradient);
                    if denom > A::zero() {
                        let scale = dot_gm / denom;
                        projected = &projected - &memory_gradient.mapv(|g| g * scale);
                    }
                }
            }
        }
        projected
    }

    /// Compute task similarity
    pub fn compute_task_similarity(&self, task1: &str, task2: &str) -> f64 {
        self.task_graph
            .task_similarities
            .get(&(task1.to_string(), task2.to_string()))
            .or_else(|| {
                self.task_graph
                    .task_similarities
                    .get(&(task2.to_string(), task1.to_string()))
            })
            .copied()
            .unwrap_or(0.0)
    }

    /// Get lifelong learning statistics
    pub fn get_lifelong_stats(&self) -> LifelongStats<A> {
        let num_tasks = self.task_optimizers.len();
        let avg_performance = if self.task_performance.is_empty() {
            A::zero()
        } else {
            let total_performance: A = self.task_performance.values().flatten().copied().sum();
            let total_samples = self
                .task_performance
                .values()
                .map(|v| v.len())
                .sum::<usize>();
            if total_samples > 0 {
                total_performance / scalar_or(total_samples, A::one())
            } else {
                A::zero()
            }
        };

        LifelongStats {
            num_tasks,
            average_performance: avg_performance,
            memory_usage: self.memory_buffer.examples.len(),
            // Mean strength of the warm starts actually performed; zero when no
            // task has been transferred into.
            transfer_efficiency: scalar_or(self.mean_transfer_weight(), A::zero()),
            catastrophic_forgetting: scalar_or(self.measured_forgetting(), A::zero()),
        }
    }

    /// Backward transfer measured from the losses recorded for tasks that are no
    /// longer current.
    ///
    /// A task's *reference* loss is the last one recorded while it was the
    /// active task. Anything recorded for it afterwards — a re-evaluation the
    /// caller performs with [`Self::record_task_performance`] — is compared
    /// against that reference, and the mean positive degradation across tasks is
    /// the forgetting measure (Lopez-Paz & Ranzato, "Gradient Episodic Memory
    /// for Continual Learning", NeurIPS 2017, report the same quantity with the
    /// opposite sign).
    ///
    /// It is `0` until an old task is actually re-evaluated: forgetting is not
    /// observable without measuring the old task again, and reporting a
    /// non-zero constant instead — which this used to do — would be a
    /// fabrication.
    fn measured_forgetting(&self) -> f64 {
        let mut total = 0.0;
        let mut counted = 0usize;

        for (task_id, reference) in &self.task_reference_loss {
            let Some(history) = self.task_performance.get(task_id) else {
                continue;
            };
            let Some(&latest) = history.last() else {
                continue;
            };
            let (Ok(latest), Ok(reference)) = (try_f64(latest), try_f64(*reference)) else {
                continue;
            };
            if !latest.is_finite() || !reference.is_finite() {
                continue;
            }
            total += (latest - reference).max(0.0);
            counted += 1;
        }

        if counted == 0 {
            0.0
        } else {
            total / counted as f64
        }
    }

    /// Record a loss measured for a task that is not the active one.
    ///
    /// This is how a caller feeds re-evaluation of previously-learned tasks back
    /// in, which is what makes [`LifelongStats::catastrophic_forgetting`]
    /// measurable. Recording against the active task is rejected: its losses
    /// arrive through [`Self::update_current_task`], and mixing the two would
    /// move the reference the measurement is taken against.
    pub fn record_task_performance(&mut self, task_id: &str, loss: A) -> Result<()> {
        if self.current_task.as_deref() == Some(task_id) {
            return Err(OptimError::InvalidConfig(format!(
                "task '{task_id}' is the active task; its losses are recorded by \
                 update_current_task"
            )));
        }
        match self.task_performance.get_mut(task_id) {
            Some(history) => {
                history.push(loss);
                Ok(())
            }
            None => Err(OptimError::InvalidConfig(format!(
                "unknown task '{task_id}'"
            ))),
        }
    }
}

/// Lifelong learning statistics
#[derive(Debug, Clone)]
pub struct LifelongStats<A: Float> {
    /// Number of tasks learned
    pub num_tasks: usize,
    /// Average performance across all tasks
    pub average_performance: A,
    /// Current memory usage
    pub memory_usage: usize,
    /// Transfer learning efficiency
    pub transfer_efficiency: A,
    /// Catastrophic forgetting measure
    pub catastrophic_forgetting: A,
}

#[cfg(test)]
mod tests {
    use super::*;
    use approx::assert_relative_eq;
    use scirs2_core::ndarray::{Array1, Ix1};

    #[test]
    fn test_online_optimizer_creation() {
        let strategy = OnlineLearningStrategy::AdaptiveSGD {
            initial_lr: 0.01,
            adaptation_method: LearningRateAdaptation::AdaGrad { epsilon: 1e-8 },
        };

        let initial_params = Array1::from_vec(vec![1.0, 2.0, 3.0]);
        let optimizer = OnlineOptimizer::new(strategy, initial_params);

        assert_eq!(optimizer.step_count, 0);
        assert_relative_eq!(optimizer.current_lr, 0.01, epsilon = 1e-6);
    }

    #[test]
    fn test_online_update() {
        let strategy = OnlineLearningStrategy::AdaptiveSGD {
            initial_lr: 0.1,
            adaptation_method: LearningRateAdaptation::ExponentialDecay { decay_rate: 0.99 },
        };

        let initial_params = Array1::from_vec(vec![1.0, 2.0, 3.0]);
        let mut optimizer = OnlineOptimizer::new(strategy, initial_params);

        let gradient = Array1::from_vec(vec![0.1, 0.2, 0.3]);
        let loss = 0.5;

        optimizer
            .online_update(&gradient, loss)
            .expect("unwrap failed");

        assert_eq!(optimizer.step_count, 1);
        assert_eq!(optimizer.performance_history.len(), 1);
        assert_relative_eq!(optimizer.performance_history[0], 0.5, epsilon = 1e-6);
    }

    #[test]
    fn test_lifelong_optimizer_creation() {
        let strategy = LifelongStrategy::ElasticWeightConsolidation {
            importance_weight: 1000.0,
            fisher_samples: 100,
        };

        let optimizer = LifelongOptimizer::<f64, scirs2_core::ndarray::Ix1>::new(strategy);

        assert_eq!(optimizer.task_optimizers.len(), 0);
        assert!(optimizer.current_task.is_none());
    }

    #[test]
    fn test_task_management() {
        let strategy = LifelongStrategy::MemoryAugmented {
            memory_size: 100,
            update_strategy: MemoryUpdateStrategy::FIFO,
        };

        let mut optimizer = LifelongOptimizer::<f64, scirs2_core::ndarray::Ix1>::new(strategy);
        let initial_params = Array1::from_vec(vec![1.0, 2.0, 3.0]);

        optimizer
            .start_task("task1".to_string(), initial_params)
            .expect("unwrap failed");

        assert_eq!(optimizer.current_task, Some("task1".to_string()));
        assert!(optimizer.task_optimizers.contains_key("task1"));
        assert!(optimizer.task_performance.contains_key("task1"));
    }

    #[test]
    fn test_memory_buffer_update() {
        let strategy = LifelongStrategy::MemoryAugmented {
            memory_size: 2,
            update_strategy: MemoryUpdateStrategy::FIFO,
        };

        let mut optimizer = LifelongOptimizer::<f64, scirs2_core::ndarray::Ix1>::new(strategy);
        optimizer.memory_buffer.max_size = 2;

        let initial_params = Array1::from_vec(vec![1.0, 2.0, 3.0]);
        optimizer
            .start_task("task1".to_string(), initial_params)
            .expect("unwrap failed");

        let gradient = Array1::from_vec(vec![0.1, 0.2, 0.3]);

        // Add first example
        optimizer
            .update_current_task(&gradient, 0.5)
            .expect("unwrap failed");
        assert_eq!(optimizer.memory_buffer.examples.len(), 1);

        // Add second example
        optimizer
            .update_current_task(&gradient, 0.6)
            .expect("unwrap failed");
        assert_eq!(optimizer.memory_buffer.examples.len(), 2);

        // Add third example (should remove first due to FIFO)
        optimizer
            .update_current_task(&gradient, 0.7)
            .expect("unwrap failed");
        assert_eq!(optimizer.memory_buffer.examples.len(), 2);
    }

    #[test]
    fn test_performance_metrics() {
        let strategy = OnlineLearningStrategy::AdaptiveSGD {
            initial_lr: 0.01,
            adaptation_method: LearningRateAdaptation::Adam {
                beta1: 0.9,
                beta2: 0.999,
                epsilon: 1e-8,
            },
        };

        let initial_params = Array1::from_vec(vec![1.0, 2.0, 3.0]);
        let mut optimizer = OnlineOptimizer::new(strategy, initial_params);

        // Add some performance data
        optimizer.performance_history.push_back(0.8);
        optimizer.performance_history.push_back(0.6);
        optimizer.performance_history.push_back(0.4);
        optimizer.regret_bound = 0.5;

        let metrics = optimizer.get_performance_metrics();

        assert_relative_eq!(metrics.cumulative_regret, 0.5, epsilon = 1e-6);
        assert_relative_eq!(metrics.average_loss, 0.6, epsilon = 1e-6);
    }

    #[test]
    fn test_lifelong_stats() {
        let strategy = LifelongStrategy::MetaLearning {
            meta_lr: 0.001,
            inner_steps: 5,
            task_embedding_size: 64,
        };

        let mut optimizer = LifelongOptimizer::<f64, scirs2_core::ndarray::Ix1>::new(strategy);

        // Add some tasks
        let initial_params = Array1::from_vec(vec![1.0, 2.0, 3.0]);
        optimizer
            .start_task("task1".to_string(), initial_params.clone())
            .expect("unwrap failed");
        optimizer
            .start_task("task2".to_string(), initial_params)
            .expect("unwrap failed");

        // Add some performance data
        optimizer
            .task_performance
            .get_mut("task1")
            .expect("unwrap failed")
            .extend(vec![0.8, 0.7]);
        optimizer
            .task_performance
            .get_mut("task2")
            .expect("unwrap failed")
            .extend(vec![0.9, 0.8]);

        let stats = optimizer.get_lifelong_stats();

        assert_eq!(stats.num_tasks, 2);
        assert_relative_eq!(stats.average_performance, 0.8, epsilon = 1e-6);
    }

    #[test]
    fn test_learning_rate_adaptations() {
        let strategies = vec![
            LearningRateAdaptation::AdaGrad { epsilon: 1e-8 },
            LearningRateAdaptation::RMSprop {
                decay: 0.9,
                epsilon: 1e-8,
            },
            LearningRateAdaptation::Adam {
                beta1: 0.9,
                beta2: 0.999,
                epsilon: 1e-8,
            },
            LearningRateAdaptation::ExponentialDecay { decay_rate: 0.99 },
            LearningRateAdaptation::InverseScaling { power: 0.5 },
        ];

        for adaptation in strategies {
            let strategy = OnlineLearningStrategy::AdaptiveSGD {
                initial_lr: 0.01,
                adaptation_method: adaptation,
            };

            let initial_params = Array1::from_vec(vec![1.0, 2.0, 3.0]);
            let mut optimizer = OnlineOptimizer::new(strategy, initial_params);

            let gradient = Array1::from_vec(vec![0.1, 0.2, 0.3]);
            let result = optimizer.online_update(&gradient, 0.5);

            assert!(result.is_ok());
            assert_eq!(optimizer.step_count, 1);
        }
    }

    /// F81: the GEM projection must remove the component of a new gradient
    /// that conflicts with a stored past-task gradient, so the previously
    /// dead replay data actually shapes the update.
    #[test]
    fn gem_projection_removes_conflicting_component() {
        let strategy = LifelongStrategy::GradientEpisodicMemory {
            memory_per_task: 10,
            violation_tolerance: 0.0,
        };
        let mut opt = LifelongOptimizer::<f64, scirs2_core::ndarray::Ix1>::new(strategy);
        let params = Array1::from_vec(vec![0.0, 0.0, 0.0]);
        opt.start_task("t".to_string(), params).expect("start_task");

        // Store a past-task gradient g1 = [1, 0, 0] in episodic memory.
        let g1 = Array1::from_vec(vec![1.0, 0.0, 0.0]);
        opt.update_current_task(&g1, 0.5).expect("update");

        // A conflicting new gradient g2 with dot(g2, g1) < 0.
        let g2 = Array1::from_vec(vec![-1.0, 1.0, 0.0]);
        let projected = opt.project_gradient_gem(&g2);

        let dot: f64 = projected.iter().zip(g1.iter()).map(|(&x, &y)| x * y).sum();
        assert!(
            dot >= -1e-9,
            "GEM projection did not remove the conflicting component (F81): dot={dot}"
        );
    }

    /// F81: `get_performance_metrics` must derive its fields from real state
    /// instead of returning hardcoded `1.0`/`step_count`/`0.8`.
    #[test]
    fn performance_metrics_are_derived_from_state() {
        let strategy = OnlineLearningStrategy::AdaptiveSGD {
            initial_lr: 0.01,
            adaptation_method: LearningRateAdaptation::AdaGrad { epsilon: 1e-8 },
        };
        let params = Array1::from_vec(vec![1.0, 2.0, 3.0]);
        let mut opt = OnlineOptimizer::new(strategy, params);

        let gradient = Array1::from_vec(vec![0.05, 0.05, 0.05]);
        let mut loss = 1.0_f64;
        for _ in 0..30 {
            opt.online_update(&gradient, loss).expect("update");
            loss *= 0.9;
        }

        let metrics = opt.get_performance_metrics();
        assert!(
            metrics.adaptation_speed > 0.0,
            "adaptation_speed not derived from the loss trend (F81): {}",
            metrics.adaptation_speed
        );
        assert!(
            metrics.lr_stability > 0.0 && metrics.lr_stability <= 1.0,
            "lr_stability out of range (F81): {}",
            metrics.lr_stability
        );
        // AdaGrad keeps no second-moment array, so memory efficiency is
        // params / (params + grad_accumulator) = 0.5, not the old 0.8.
        assert!(
            (metrics.memory_efficiency - 0.5).abs() < 1e-9,
            "memory_efficiency not derived from real optimizer state (F81): {}",
            metrics.memory_efficiency
        );
    }

    // --- Lifelong-strategy wiring regression tests -----------------------
    //
    // `apply_ewc_regularization`, `apply_progressive_networks` and
    // `apply_meta_learning` were all `Ok(())` no-ops that ignored their
    // `gradient` argument, and `SharedKnowledge::{fisher_information,
    // important_parameters, meta_parameters}` were `#[allow(dead_code)]`
    // fields nothing ever wrote. These tests pin the real behaviour.

    fn ewc_optimizer(importance_weight: f64) -> LifelongOptimizer<f64, Ix1> {
        LifelongOptimizer::new(LifelongStrategy::ElasticWeightConsolidation {
            importance_weight,
            fisher_samples: 100,
        })
    }

    /// The diagonal empirical Fisher must actually be accumulated from the
    /// observed gradients (it stayed `None` forever before).
    #[test]
    fn ewc_accumulates_diagonal_fisher_information() {
        let mut opt = ewc_optimizer(1.0);
        opt.start_task("a".to_string(), Array1::zeros(3))
            .expect("start_task");
        assert!(
            opt.fisher_information().is_none(),
            "no gradient observed yet, so there can be no Fisher estimate"
        );

        let gradient = Array1::from_vec(vec![2.0, 0.0, -4.0]);
        for _ in 0..5 {
            opt.update_current_task(&gradient, 1.0)
                .expect("update_current_task");
        }

        let fisher = opt
            .fisher_information()
            .expect("Fisher information must exist after observing gradients");
        // A constant gradient g gives a running mean of exactly g^2.
        assert!((fisher[0] - 4.0).abs() < 1e-9, "F[0] = {}", fisher[0]);
        assert!((fisher[1] - 0.0).abs() < 1e-9, "F[1] = {}", fisher[1]);
        assert!((fisher[2] - 16.0).abs() < 1e-9, "F[2] = {}", fisher[2]);
    }

    /// Switching tasks must consolidate the outgoing task's parameters into
    /// the shared EWC anchor `theta*`.
    #[test]
    fn starting_a_new_task_consolidates_the_previous_parameters() {
        let mut opt = ewc_optimizer(1.0);
        opt.start_task("a".to_string(), Array1::from_vec(vec![1.0, 2.0]))
            .expect("start_task a");
        assert!(
            opt.consolidated_parameters().is_none(),
            "nothing has been consolidated before the first task switch"
        );

        opt.start_task("b".to_string(), Array1::zeros(2))
            .expect("start_task b");
        let anchor = opt
            .consolidated_parameters()
            .expect("task a must have been consolidated on switching to b");
        assert_eq!(anchor.len(), 2);
    }

    /// The EWC penalty must pull the *new* task's parameters back toward the
    /// consolidated anchor: with a large importance weight, a task whose own
    /// gradient is zero must still move toward `theta*` instead of standing
    /// still (which is what the old no-op produced).
    #[test]
    fn ewc_penalty_pulls_parameters_toward_the_consolidated_anchor() {
        let mut opt = ewc_optimizer(1000.0);

        // Task A settles at a non-zero parameter vector, and produces a
        // gradient so a Fisher estimate exists for those coordinates.
        opt.start_task("a".to_string(), Array1::from_vec(vec![5.0, 5.0]))
            .expect("start_task a");
        let probe = Array1::from_vec(vec![1.0, 1.0]);
        opt.update_current_task(&probe, 1.0).expect("task a step");
        let anchor = opt
            .task_optimizers
            .get("a")
            .expect("task a optimizer")
            .parameters()
            .clone();

        // Task B starts far away from the anchor with a zero task gradient:
        // the only force acting on it is the EWC penalty.
        opt.start_task("b".to_string(), Array1::from_vec(vec![50.0, 50.0]))
            .expect("start_task b");
        let before = opt
            .task_optimizers
            .get("b")
            .expect("task b optimizer")
            .parameters()
            .clone();
        let zero = Array1::zeros(2);
        opt.update_current_task(&zero, 1.0).expect("task b step");
        let after = opt
            .task_optimizers
            .get("b")
            .expect("task b optimizer")
            .parameters()
            .clone();

        assert_ne!(
            before, after,
            "EWC regression: a zero task gradient left the parameters \
             untouched, so the penalty term is not being applied"
        );
        let moved_toward_anchor = (after[0] - anchor[0]).abs() < (before[0] - anchor[0]).abs();
        assert!(
            moved_toward_anchor,
            "EWC penalty moved the parameter away from the anchor \
             (anchor={}, before={}, after={})",
            anchor[0], before[0], after[0]
        );
    }

    /// A zero importance weight must disable the penalty entirely, so the
    /// weight is genuinely read rather than ignored.
    #[test]
    fn zero_importance_weight_disables_the_ewc_penalty() {
        let mut opt = ewc_optimizer(0.0);
        opt.start_task("a".to_string(), Array1::from_vec(vec![5.0, 5.0]))
            .expect("start_task a");
        opt.update_current_task(&Array1::from_vec(vec![1.0, 1.0]), 1.0)
            .expect("task a step");
        opt.start_task("b".to_string(), Array1::from_vec(vec![50.0, 50.0]))
            .expect("start_task b");

        let before = opt
            .task_optimizers
            .get("b")
            .expect("task b optimizer")
            .parameters()
            .clone();
        opt.update_current_task(&Array1::zeros(2), 1.0)
            .expect("task b step");
        let after = opt
            .task_optimizers
            .get("b")
            .expect("task b optimizer")
            .parameters()
            .clone();

        assert_eq!(
            before, after,
            "with importance_weight = 0 the EWC penalty must vanish"
        );
    }

    /// Reptile must move the shared meta-parameters toward the adapted task
    /// parameters; they used to stay `None` forever.
    #[test]
    fn meta_learning_maintains_reptile_meta_parameters() {
        let mut opt: LifelongOptimizer<f64, Ix1> =
            LifelongOptimizer::new(LifelongStrategy::MetaLearning {
                meta_lr: 0.5,
                inner_steps: 1,
                task_embedding_size: 4,
            });
        assert!(opt.meta_parameters().is_none());

        opt.start_task("a".to_string(), Array1::from_vec(vec![0.0, 0.0]))
            .expect("start_task a");
        let gradient = Array1::from_vec(vec![1.0, -1.0]);
        opt.update_current_task(&gradient, 1.0)
            .expect("task a step");
        let seeded = opt
            .meta_parameters()
            .expect("meta-parameters must be seeded from the first task")
            .clone();

        // A second task that starts far away must drag the meta-parameters
        // toward it by exactly meta_lr of the gap.
        opt.start_task("b".to_string(), Array1::from_vec(vec![10.0, 10.0]))
            .expect("start_task b");
        opt.update_current_task(&Array1::zeros(2), 1.0)
            .expect("task b step");
        let updated = opt.meta_parameters().expect("meta-parameters").clone();
        let task_b = opt
            .task_optimizers
            .get("b")
            .expect("task b optimizer")
            .parameters()
            .clone();

        let expected = seeded[0] + (task_b[0] - seeded[0]) * 0.5;
        assert!(
            (updated[0] - expected).abs() < 1e-9,
            "Reptile update wrong: seeded={}, task={}, expected={}, got={}",
            seeded[0],
            task_b[0],
            expected,
            updated[0]
        );
    }

    /// Progressive Networks is not implementable without a model-column
    /// partition, so it must report that honestly instead of silently running
    /// plain online SGD and returning `Ok(())`.
    #[test]
    fn progressive_networks_reports_that_it_is_unsupported() {
        let mut opt: LifelongOptimizer<f64, Ix1> =
            LifelongOptimizer::new(LifelongStrategy::ProgressiveNetworks {
                lateral_strength: 0.5,
                growth_strategy: ColumnGrowthStrategy::PerTask,
            });
        opt.start_task("a".to_string(), Array1::zeros(2))
            .expect("start_task");
        let err = opt
            .update_current_task(&Array1::from_vec(vec![1.0, 1.0]), 1.0)
            .expect_err("ProgressiveNetworks must not fabricate success");
        assert!(
            matches!(err, OptimError::UnsupportedOperation(_)),
            "expected UnsupportedOperation, got {err:?}"
        );
    }
}