foxhunt/ml/tests/dqn_hyperparameter_test.rs

//! DQN Hyperparameter Tests
//!
//! Test suite for DQN hyperparameter presets, validation, and CLI configuration.
//!
//! Coverage:
//! - Test 1: Conservative preset values
//! - Test 2: Aggressive preset values (to be implemented)
//! - Test 3: Production preset values (Trial #35 optimal)
//! - Test 4: CLI configurability of all hyperparameters
//! - Test 5: Hyperopt adapter uses correct ranges
//! - Test 6: Hyperparameter validation (reject invalid values)

use ml::hyperopt::adapters::dqn::{DQNParams, DQNTrainer};
use ml::hyperopt::traits::ParameterSpace;
use ml::trainers::dqn::DQNHyperparameters;

/// Test 1: Verify conservative() preset has correct values
///
/// Conservative preset should be suitable for testing and development.
#[test]
fn test_conservative_preset() {
    let params = DQNHyperparameters::conservative();

    // Learning rate: moderate value
    assert_eq!(params.learning_rate, 0.0001);

    // Batch size: medium value (GPU compatible)
    assert_eq!(params.batch_size, 128);

    // Gamma: high value for long-term rewards
    assert_eq!(params.gamma, 0.99);

    // Epsilon parameters: standard exploration schedule
    assert_eq!(params.epsilon_start, 1.0);
    assert_eq!(params.epsilon_end, 0.01);
    assert_eq!(params.epsilon_decay, 0.995);

    // Buffer size: moderate capacity
    assert_eq!(params.buffer_size, 100_000);
    assert_eq!(params.min_replay_size, 1_000);

    // Training config
    assert_eq!(params.epochs, 100);
    assert_eq!(params.checkpoint_frequency, 10);

    // Early stopping: enabled with conservative thresholds
    assert!(params.early_stopping_enabled);
    assert_eq!(params.q_value_floor, 0.5);
    assert_eq!(params.min_loss_improvement_pct, 2.0);
    assert_eq!(params.plateau_window, 30);
    assert_eq!(params.min_epochs_before_stopping, 50);

    // Advanced features: enabled
    assert_eq!(params.gradient_clip_norm, Some(1.0));
    assert!(params.use_huber_loss);
    assert_eq!(params.huber_delta, 1.0);
    assert!(params.use_double_dqn);

    // HOLD penalty
    assert_eq!(params.hold_penalty_weight, 0.01);
    // Target network updates
    assert_eq!(params.target_update_frequency, 500);
    assert_eq!(params.target_update_tau, None);

    // Validation configuration
    assert!(!params.skip_validation);
    assert_eq!(params.validation_split, 0.2);
    assert_eq!(params.validation_log_frequency, 1);

    assert_eq!(params.movement_threshold, 0.02);
}

/// Test 2: Verify aggressive() preset has correct values
///
/// Aggressive preset should use higher learning rates and larger batch sizes
/// for faster convergence (at the risk of instability).
#[test]
fn test_aggressive_preset() {
    let params = DQNHyperparameters::aggressive();

    // Learning rate: higher for faster learning
    assert_eq!(params.learning_rate, 0.0005);

    // Batch size: larger for more stable gradients
    assert_eq!(params.batch_size, 230); // Max for RTX 3050 Ti

    // Gamma: lower for more short-term focus
    assert_eq!(params.gamma, 0.95);

    // Epsilon: faster decay for exploitation
    assert_eq!(params.epsilon_start, 1.0);
    assert_eq!(params.epsilon_end, 0.01);
    assert_eq!(params.epsilon_decay, 0.99); // Faster decay than conservative

    // Buffer size: larger for more diverse experiences
    assert_eq!(params.buffer_size, 200_000);
    assert_eq!(params.min_replay_size, 2_000);

    // Training config
    assert_eq!(params.epochs, 500);
    assert_eq!(params.checkpoint_frequency, 50);

    // Early stopping: disabled for aggressive training
    assert!(!params.early_stopping_enabled);

    // Advanced features
    assert_eq!(params.gradient_clip_norm, Some(1.0));
    // Target network updates
    assert_eq!(params.target_update_frequency, 1000);
    assert_eq!(params.target_update_tau, Some(0.005));

    assert!(params.use_huber_loss);
    assert!(params.use_double_dqn);
}

/// Test 3: Verify production() preset has optimal Trial #35 values
///
/// Production preset should use hyperopt-optimized values from Trial #35:
/// - Learning rate: 5e-6 (more stable than 1e-5)
/// - Batch size: 64 (better than 110)
/// - Gamma: 0.92 (less long-term bias than 0.9775)
/// - Epsilon decay: 0.997 (slower than 0.9394)
/// - Buffer size: 50,000 (larger than 34,000)
#[test]
fn test_production_preset() {
    let params = DQNHyperparameters::production();

    // Optimized hyperparameters from Trial #35 backtesting analysis
    assert_eq!(params.learning_rate, 5e-6);
    assert_eq!(params.batch_size, 64);
    assert_eq!(params.gamma, 0.92);

    // Epsilon schedule: slower decay for more exploration
    assert_eq!(params.epsilon_start, 1.0);
    assert_eq!(params.epsilon_end, 0.01);
    assert_eq!(params.epsilon_decay, 0.997);

    // Buffer size: larger for better experience diversity
    assert_eq!(params.buffer_size, 50_000);
    assert_eq!(params.min_replay_size, 128); // 2x batch size

    // Training config: production defaults
    assert_eq!(params.epochs, 1000); // Longer training for production
    assert_eq!(params.checkpoint_frequency, 100);

    // Early stopping: enabled with production thresholds
    assert!(params.early_stopping_enabled);
    assert_eq!(params.q_value_floor, 0.5);
    assert_eq!(params.min_loss_improvement_pct, 0.1); // More sensitive
    assert_eq!(params.plateau_window, 5);
    assert_eq!(params.min_epochs_before_stopping, 50);

    // Advanced features: all enabled for production
    assert_eq!(params.gradient_clip_norm, Some(1.0));
    assert!(params.use_huber_loss);
    assert_eq!(params.huber_delta, 1.0);
    assert!(params.use_double_dqn);

    // HOLD penalty: default values
    assert_eq!(params.hold_penalty_weight, 0.01);
    assert_eq!(params.movement_threshold, 0.02);
}

/// Test 4: Verify all hyperparameters are accessible and configurable
///
/// This test ensures all DQNHyperparameters fields can be set and retrieved.
#[test]
// Target network updates
assert_eq!(params.target_update_frequency, 500);
assert_eq!(params.target_update_tau, None);

// Validation configuration
assert!(!params.skip_validation);
assert_eq!(params.validation_split, 0.2);
assert_eq!(params.validation_log_frequency, 1);

fn test_all_hyperparameters_configurable() {
    // Create custom hyperparameters by modifying production preset
    let mut params = DQNHyperparameters::production();

    // Verify all fields are mutable and accessible
    params.learning_rate = 1e-4;
    params.batch_size = 128;
    params.gamma = 0.99;
    params.epsilon_start = 0.5;
    params.epsilon_end = 0.05;
    params.epsilon_decay = 0.995;
    params.buffer_size = 100_000;
    params.min_replay_size = 500;
    params.epochs = 200;
    params.checkpoint_frequency = 20;
    params.early_stopping_enabled = false;
    params.q_value_floor = 1.0;
    params.min_loss_improvement_pct = 5.0;
    params.plateau_window = 10;
    params.min_epochs_before_stopping = 100;
    params.hold_penalty_weight = 0.02;
    params.movement_threshold = 0.01;
    params.gradient_clip_norm = Some(0.5);
    params.use_huber_loss = false;
    params.huber_delta = 2.0;
    params.use_double_dqn = false;

    // Verify values were set correctly
    assert_eq!(params.learning_rate, 1e-4);
    assert_eq!(params.batch_size, 128);
    assert_eq!(params.gamma, 0.99);
    assert_eq!(params.epsilon_start, 0.5);
    assert_eq!(params.epsilon_end, 0.05);
    assert_eq!(params.epsilon_decay, 0.995);
    assert_eq!(params.buffer_size, 100_000);
    assert_eq!(params.min_replay_size, 500);
    assert_eq!(params.epochs, 200);
    assert_eq!(params.checkpoint_frequency, 20);
    assert!(!params.early_stopping_enabled);
    assert_eq!(params.q_value_floor, 1.0);
    assert_eq!(params.min_loss_improvement_pct, 5.0);
    assert_eq!(params.plateau_window, 10);
    assert_eq!(params.min_epochs_before_stopping, 100);
    assert_eq!(params.hold_penalty_weight, 0.02);
    assert_eq!(params.movement_threshold, 0.01);
    assert_eq!(params.gradient_clip_norm, Some(0.5));
    assert!(!params.use_huber_loss);
    assert_eq!(params.huber_delta, 2.0);
    assert!(!params.use_double_dqn);
}

/// Test 5: Verify hyperopt adapter uses correct parameter ranges
///
/// Updated ranges based on Trial #35 backtesting analysis:
/// - Learning rate: 1e-6 to 1e-4 (narrower, more conservative)
/// - Gamma: 0.90 to 0.95 (lower values, less long-term bias)
/// - Epsilon decay: 0.995 to 0.9995 (slower decay)
/// - Batch size: 32 to 128 (GPU optimized)
/// - Buffer size: 10,000 to 100,000 (expanded range)
#[test]
fn test_hyperopt_ranges() {
    let bounds = DQNParams::continuous_bounds();

    // 5 continuous parameters: [learning_rate, batch_size, gamma, epsilon_decay, buffer_size]
    assert_eq!(bounds.len(), 5);

    // Learning rate: log-scale, 1e-6 to 1e-4 (updated from 1e-5 to 1e-3)
    let lr_bounds = bounds[0];
    assert!((lr_bounds.0 - (1e-6_f64).ln()).abs() < 1e-10);
    assert!((lr_bounds.1 - (1e-4_f64).ln()).abs() < 1e-10);

    // Batch size: linear scale, 32 to 128 (updated from 64 to 230)
    let batch_bounds = bounds[1];
    assert_eq!(batch_bounds, (32.0, 128.0));

    // Gamma: linear scale, 0.90 to 0.95 (updated from 0.95 to 0.99)
    let gamma_bounds = bounds[2];
    assert_eq!(gamma_bounds, (0.90, 0.95));

    // Epsilon decay: log-scale, 0.995 to 0.9995 (updated from 0.88 to 0.95)
    let epsilon_bounds = bounds[3];
    assert!((epsilon_bounds.0 - (0.995_f64).ln()).abs() < 1e-10);
    assert!((epsilon_bounds.1 - (0.9995_f64).ln()).abs() < 1e-10);

    // Buffer size: log-scale, 10k to 100k (updated from 10k to 1M)
    let buffer_bounds = bounds[4];
    assert!((buffer_bounds.0 - (10_000_f64).ln()).abs() < 1e-10);
    assert!((buffer_bounds.1 - (100_000_f64).ln()).abs() < 1e-10);
}

/// Test 6: Verify hyperparameter validation rejects invalid values
///
/// Validation should catch:
/// - Learning rate outside (0, 1]
/// - Gamma outside (0, 1)
/// - Epsilon decay outside (0, 1)
/// - Batch size < 1
/// - Buffer size < batch size
#[test]
fn test_hyperparameter_validation() {
    // Valid production parameters should pass
    let valid_params = DQNHyperparameters::production();
    assert!(valid_params.validate().is_ok());

    // Test 1: Invalid learning rate (too low)
    let mut invalid_lr_low = valid_params.clone();
    invalid_lr_low.learning_rate = 0.0;
    assert!(invalid_lr_low.validate().is_err());
    assert_eq!(
        invalid_lr_low.validate().unwrap_err(),
        "learning_rate must be in (0, 1]"
    );

    // Test 2: Invalid learning rate (too high)
    let mut invalid_lr_high = valid_params.clone();
    invalid_lr_high.learning_rate = 1.5;
    assert!(invalid_lr_high.validate().is_err());
    assert_eq!(
        invalid_lr_high.validate().unwrap_err(),
        "learning_rate must be in (0, 1]"
    );

    // Test 3: Invalid gamma (too low)
    let mut invalid_gamma_low = valid_params.clone();
    invalid_gamma_low.gamma = 0.0;
    assert!(invalid_gamma_low.validate().is_err());
    assert_eq!(
        invalid_gamma_low.validate().unwrap_err(),
        "gamma must be in (0, 1)"
    );

    // Test 4: Invalid gamma (too high)
    let mut invalid_gamma_high = valid_params.clone();
    invalid_gamma_high.gamma = 1.0;
    assert!(invalid_gamma_high.validate().is_err());
    assert_eq!(
        invalid_gamma_high.validate().unwrap_err(),
        "gamma must be in (0, 1)"
    );

    // Test 5: Invalid epsilon_decay (too low)
    let mut invalid_epsilon_low = valid_params.clone();
    invalid_epsilon_low.epsilon_decay = 0.0;
    assert!(invalid_epsilon_low.validate().is_err());
    assert_eq!(
        invalid_epsilon_low.validate().unwrap_err(),
        "epsilon_decay must be in (0, 1)"
    );

    // Test 6: Invalid epsilon_decay (too high)
    let mut invalid_epsilon_high = valid_params.clone();
    invalid_epsilon_high.epsilon_decay = 1.0;
    assert!(invalid_epsilon_high.validate().is_err());
    assert_eq!(
        invalid_epsilon_high.validate().unwrap_err(),
        "epsilon_decay must be in (0, 1)"
    );

    // Test 7: Invalid batch size (< 1)
    let mut invalid_batch = valid_params.clone();
    invalid_batch.batch_size = 0;
    assert!(invalid_batch.validate().is_err());
    assert_eq!(
        invalid_batch.validate().unwrap_err(),
        "batch_size must be >= 1"
    );

    // Test 8: Invalid buffer size (< batch size)
    let mut invalid_buffer = valid_params.clone();
    invalid_buffer.buffer_size = invalid_buffer.batch_size - 1;
    assert!(invalid_buffer.validate().is_err());
    assert_eq!(
        invalid_buffer.validate().unwrap_err(),
        "buffer_size must be >= batch_size"
    );

    // Test 9: Invalid target_update_frequency (< 1)
    let mut invalid_target_freq = valid_params.clone();
    invalid_target_freq.target_update_frequency = 0;
    assert!(invalid_target_freq.validate().is_err());
    assert_eq!(
        invalid_target_freq.validate().unwrap_err(),
        "target_update_frequency must be >= 1"
    );
}

/// Test 7: Verify parameter space conversions are accurate
///
/// Test that DQNParams can correctly convert to/from continuous representation
/// and that bounds are enforced.
#[test]
fn test_parameter_space_conversions() {
    // Create params with known values
    let original = DQNParams {
        learning_rate: 5e-6,
        batch_size: 64,
        gamma: 0.92,
        epsilon_decay: 0.997,
        buffer_size: 50_000,
    };

    // Convert to continuous and back
    let continuous = original.to_continuous();
    let recovered = DQNParams::from_continuous(&continuous).unwrap();

    // Verify values match (with small floating point tolerance)
    assert!((recovered.learning_rate - original.learning_rate).abs() < 1e-10);
    assert_eq!(recovered.batch_size, original.batch_size);
    assert!((recovered.gamma - original.gamma).abs() < 1e-10);
    assert!((recovered.epsilon_decay - original.epsilon_decay).abs() < 1e-6);
    assert_eq!(recovered.buffer_size, original.buffer_size);
}