c645e6222d
CRITICAL FINDINGS from 3-trial validation: - 85,120 gradient clipping warnings (81.6% of logs) - REGRESSION - Rainbow features DISABLED: use_dueling=false, use_distributional=false, use_noisy_nets=false - Negative Q-values confirmed: HOLD -1000 to -3250 - Performance: Sharpe 0.29 (target 0.77) Changes: - Fixed N-Step compilation (7/7 tests passing) - Fixed Distributional compilation (6/6 tests passing) - Fixed Dueling CUDA errors (10/10 tests passing) - Added TDD validation for state_dim=225 - Total: 23/23 Wave 11 tests passing (100%) Issues requiring investigation: 1. Why are Dueling/Distributional/Noisy disabled in hyperopt? 2. Why gradient explosion despite previous fixes? 3. Test coverage gaps - unit tests pass but integration fails 🤖 Generated with Claude Code Co-Authored-By: Claude <noreply@anthropic.com>
88 lines
2.6 KiB
Rust
88 lines
2.6 KiB
Rust
//! Test suite for PPO RewardNormalizer
|
|
//!
|
|
//! TDD Red Phase: Tests written FIRST before implementation
|
|
//! These tests define the expected behavior of the RewardNormalizer module
|
|
|
|
// Import from PPO module (will fail until implementation exists)
|
|
use ml::ppo::reward_normalizer::RewardNormalizer;
|
|
|
|
#[test]
|
|
fn test_reward_normalizer_welford() {
|
|
// Test Welford's algorithm correctness with known statistics
|
|
let mut normalizer = RewardNormalizer::new();
|
|
|
|
// Add values: [10, 20, 30, 40, 50]
|
|
let values = vec![10.0, 20.0, 30.0, 40.0, 50.0];
|
|
for &value in &values {
|
|
normalizer.update(value);
|
|
}
|
|
|
|
// Expected mean: 30.0
|
|
// Expected variance: 200.0 (population variance)
|
|
// Expected std: 14.142
|
|
let (mean, std) = normalizer.get_stats();
|
|
assert!((mean - 30.0).abs() < 1e-6, "Mean should be 30.0, got {}", mean);
|
|
assert!(
|
|
(std - 14.142135623730951).abs() < 1e-6,
|
|
"Std should be ~14.142, got {}",
|
|
std
|
|
);
|
|
|
|
// Test normalization: (40 - 30) / 14.142 ≈ 0.707
|
|
let normalized = normalizer.normalize(40.0);
|
|
assert!(
|
|
(normalized - 0.7071067811865475).abs() < 1e-6,
|
|
"Normalized value should be ~0.707, got {}",
|
|
normalized
|
|
);
|
|
|
|
// Verify count
|
|
assert_eq!(normalizer.count(), 5, "Count should be 5");
|
|
}
|
|
|
|
#[test]
|
|
fn test_reward_normalizer_numerical_stability() {
|
|
// Test with extreme values to ensure numerical stability
|
|
let mut normalizer = RewardNormalizer::new();
|
|
|
|
// Add extreme positive and negative values
|
|
let extreme_values = vec![-1000.0, -500.0, 0.0, 500.0, 1000.0];
|
|
for &value in &extreme_values {
|
|
normalizer.update(value);
|
|
}
|
|
|
|
let (mean, std) = normalizer.get_stats();
|
|
|
|
// Mean should be 0.0 (symmetric distribution)
|
|
assert!(
|
|
mean.abs() < 1e-6,
|
|
"Mean should be ~0.0 for symmetric values, got {}",
|
|
mean
|
|
);
|
|
|
|
// Std should be ~707.1 (for uniform extreme distribution)
|
|
assert!(
|
|
std > 700.0 && std < 710.0,
|
|
"Std should be ~707.1, got {}",
|
|
std
|
|
);
|
|
|
|
// Normalize a new extreme value
|
|
let normalized_extreme = normalizer.normalize(1000.0);
|
|
|
|
// Should be bounded: 1000 is 1 std above mean
|
|
assert!(
|
|
normalized_extreme > 1.0 && normalized_extreme < 2.0,
|
|
"Normalized extreme should be ~1.58, got {}",
|
|
normalized_extreme
|
|
);
|
|
|
|
// Test with very small epsilon (no division by zero)
|
|
normalizer.update(0.0);
|
|
let normalized_zero = normalizer.normalize(0.0);
|
|
assert!(
|
|
!normalized_zero.is_nan() && !normalized_zero.is_infinite(),
|
|
"Normalized value should not be NaN or Inf"
|
|
);
|
|
}
|