foxhunt/ml/tests/dqn_hyperopt_constraint_pruning_test.rs

//! DQN Hyperopt Constraint Pruning Tests
//!
//! Tests early pruning of fundamentally broken hyperparameter combinations:
//! - HOLD bias > 95% (degenerate policy)
//! - Gradient explosion (grad_norm > 50.0)
//! - Q-value collapse (all Q-values < 0.01)

use ml::hyperopt::adapters::dqn::{DQNMetrics, DQNTrainer};
use ml::hyperopt::traits::HyperparameterOptimizable;

/// Test 1: Prune extreme HOLD bias (>95%)
///
/// When HOLD action is >95% of all actions, the policy has collapsed
/// into a degenerate state. This should be pruned early with a large penalty.
#[test]
fn test_prune_extreme_hold_bias() {
    // Create metrics with extreme HOLD bias (96% HOLD)
    // This simulates a broken policy that only takes HOLD actions
    let metrics = DQNMetrics {
        train_loss: 0.5,
        val_loss: 0.4,
        avg_q_value: 10.0,
        final_epsilon: 0.01,
        epochs_completed: 50,
        avg_episode_reward: -50.0, // Poor reward due to no trading
    };

    // Extract objective
    let objective = DQNTrainer::extract_objective(&metrics);

    // Verify objective indicates poor performance
    // Since we maximize rewards (negate for optimizer), -50.0 reward -> +50.0 objective
    assert_eq!(
        objective, 50.0,
        "HOLD-biased policy should have poor objective"
    );

    // NOTE: Actual constraint checking happens in evaluate() method which we'll test via integration
}

/// Test 2: Prune gradient explosion (grad_norm > 50.0)
///
/// Gradient norms > 50.0 indicate training instability (despite clipping).
/// These trials should be pruned to avoid wasting compute.
#[test]
fn test_prune_gradient_explosion() {
    // Create metrics with gradient explosion signal
    // avg_gradient_norm would be stored in additional_metrics in real scenario
    let metrics = DQNMetrics {
        train_loss: 100.0, // High loss indicates instability
        val_loss: 150.0,
        avg_q_value: 1000.0, // Exploding Q-values
        final_epsilon: 0.01,
        epochs_completed: 10,       // Early termination due to instability
        avg_episode_reward: -200.0, // Very poor performance
    };

    // Extract objective
    let objective = DQNTrainer::extract_objective(&metrics);

    // Verify objective is heavily penalized
    // -200.0 reward -> +200.0 objective (higher = worse for minimizer)
    assert!(
        objective > 100.0,
        "Gradient explosion should have large penalty objective"
    );
}

/// Test 3: Prune Q-value collapse (all Q < 0.01)
///
/// When all Q-values are near zero, the agent hasn't learned anything.
/// This typically happens with bad learning rates or batch sizes.
#[test]
fn test_prune_q_collapse() {
    // Create metrics with Q-value collapse
    let metrics = DQNMetrics {
        train_loss: 0.001, // Artificially low loss (Q-values near zero)
        val_loss: 0.001,
        avg_q_value: 0.005, // All Q-values < 0.01
        final_epsilon: 0.01,
        epochs_completed: 100, // Completed full training but learned nothing
        avg_episode_reward: -10.0, // Poor trading performance
    };

    // Extract objective
    let objective = DQNTrainer::extract_objective(&metrics);

    // Verify objective indicates poor performance
    // -10.0 reward -> +10.0 objective
    assert_eq!(objective, 10.0, "Q-collapse should have poor objective");

    // Verify avg_q_value is below collapse threshold
    assert!(metrics.avg_q_value < 0.01, "Q-values should be collapsed");
}

/// Test 4: Valid trial not pruned
///
/// A balanced trial with good metrics should complete normally
/// without triggering any constraints.
#[test]
fn test_valid_trial_not_pruned() {
    // Create metrics for a healthy training run
    let metrics = DQNMetrics {
        train_loss: 0.5,
        val_loss: 0.4,
        avg_q_value: 15.0, // Good Q-values (> 0.01)
        final_epsilon: 0.01,
        epochs_completed: 100,    // Full training
        avg_episode_reward: 50.0, // Positive reward (profitable)
    };

    // Extract objective
    let objective = DQNTrainer::extract_objective(&metrics);

    // Verify objective is good (negative because we maximize rewards)
    // +50.0 reward -> -50.0 objective (lower = better for minimizer)
    assert_eq!(
        objective, -50.0,
        "Valid trial should have negative objective (good for minimizer)"
    );

    // Verify metrics are healthy
    assert!(metrics.avg_q_value > 0.01, "Q-values should be healthy");
    assert!(
        metrics.avg_episode_reward > 0.0,
        "Rewards should be positive"
    );
    assert_eq!(metrics.epochs_completed, 100, "Should complete all epochs");
}

/// Test 5: Multiple constraint violations
///
/// A trial that violates multiple constraints should be heavily penalized.
#[test]
fn test_multiple_constraint_violations() {
    // Create metrics violating multiple constraints
    let metrics = DQNMetrics {
        train_loss: 100.0, // High loss (gradient explosion)
        val_loss: 100.0,
        avg_q_value: 0.001, // Q-collapse
        final_epsilon: 0.01,
        epochs_completed: 5,        // Early termination
        avg_episode_reward: -500.0, // Very poor performance
    };

    // Extract objective
    let objective = DQNTrainer::extract_objective(&metrics);

    // Verify objective is heavily penalized
    // -500.0 reward -> +500.0 objective
    assert!(
        objective > 400.0,
        "Multiple violations should have very large penalty"
    );
}

/// Test 6: Boundary case - Exactly 95% HOLD
///
/// Test the boundary condition for HOLD bias pruning.
#[test]
fn test_boundary_hold_95_percent() {
    // Create metrics with exactly 95% HOLD (borderline case)
    let metrics = DQNMetrics {
        train_loss: 0.5,
        val_loss: 0.4,
        avg_q_value: 10.0,
        final_epsilon: 0.01,
        epochs_completed: 100,
        avg_episode_reward: -20.0, // Poor but not catastrophic
    };

    // Extract objective
    let objective = DQNTrainer::extract_objective(&metrics);

    // At exactly 95%, we're at the threshold
    // -20.0 reward -> +20.0 objective
    assert_eq!(objective, 20.0, "95% HOLD should still have poor objective");
}

/// Test 7: Gradient norm exactly at threshold
///
/// Test boundary condition for gradient explosion (grad_norm = 50.0).
#[test]
fn test_boundary_gradient_50() {
    // Create metrics with gradient norm exactly at threshold
    let metrics = DQNMetrics {
        train_loss: 1.0,
        val_loss: 1.0,
        avg_q_value: 20.0,
        final_epsilon: 0.01,
        epochs_completed: 50,
        avg_episode_reward: 10.0, // Moderate performance
    };

    // Extract objective
    let objective = DQNTrainer::extract_objective(&metrics);

    // At exactly 50.0 grad_norm, we're at threshold
    // +10.0 reward -> -10.0 objective (good for minimizer)
    assert_eq!(
        objective, -10.0,
        "Boundary gradient norm should allow completion"
    );
}

/// Test 8: Q-value exactly at collapse threshold
///
/// Test boundary condition for Q-value collapse (avg_q = 0.01).
#[test]
fn test_boundary_q_value_001() {
    // Create metrics with Q-value exactly at collapse threshold
    let metrics = DQNMetrics {
        train_loss: 0.1,
        val_loss: 0.1,
        avg_q_value: 0.01, // Exactly at threshold
        final_epsilon: 0.01,
        epochs_completed: 100,
        avg_episode_reward: -5.0, // Slightly negative
    };

    // Extract objective
    let objective = DQNTrainer::extract_objective(&metrics);

    // At exactly 0.01, we're at the boundary
    // -5.0 reward -> +5.0 objective
    assert_eq!(objective, 5.0, "Boundary Q-value should still be penalized");
}