BREAKING CHANGES: - Removed orphaned dqn.rs monolithic trainer (4,975 lines) - Removed orphaned dqn_ensemble.rs module (816 lines) - Removed orphaned tft.rs and tft_complete_int8_integration_test.rs - TFT trainer split into modular directory structure DQN Module Refactoring: - Split trainers/dqn.rs into modular structure (config.rs, statistics.rs, trainer.rs) - Fixed hyperopt 39D search space (continuous params only) - Boolean flags (use_dueling, use_double_dqn, use_per, use_noisy_nets) are now FIXED architectural decisions - use_distributional defaults to false (Candle BUG #36 - scatter_add gradient issues) Clean Module Structure: - ml/src/trainers/dqn/ directory with proper mod.rs exports - ml/src/trainers/tft/ directory with config.rs, types.rs, model.rs, trainer.rs, tests.rs - All P0 features validated: TD-error clamping, batch diversity, LR scheduler, priority staleness Documentation: - Added comprehensive docs in docs/codebase-cleanup/ - ADR-001 for DQN refactoring decisions - Rainbow DQN component matrix and quick reference guides Build Status: Compiles with zero errors 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
4.2 KiB
4.2 KiB
WAVE 26 P0.2: Batch Diversity Enforcement Implementation
Overview
Added batch diversity enforcement to prevent sampling the same experience twice in consecutive batches.
Implementation Changes
1. Added HashSet for Tracking Recently Sampled Indices
File: /home/jgrusewski/Work/foxhunt/ml/src/dqn/prioritized_replay.rs
Imports
use std::collections::HashSet;
Struct Fields
pub struct PrioritizedReplayBuffer {
// ... existing fields ...
/// Tracks recently sampled indices to prevent duplicates
recently_sampled: Arc<Mutex<HashSet<usize>>>,
/// Counter for cooldown management (cleared every 50 batches)
sample_counter: AtomicUsize,
}
Constructor
pub fn new(config: PrioritizedReplayConfig) -> Result<Self, MLError> {
// ... existing initialization ...
Ok(Self {
// ... existing fields ...
recently_sampled: Arc::new(Mutex::new(HashSet::new())),
sample_counter: AtomicUsize::new(0),
config,
})
}
2. Modified sample() Method
Rejection Sampling with Fallback:
pub fn sample(&self, batch_size: usize) -> Result<(Vec<Experience>, Vec<f32>, Vec<usize>), MLError> {
// ... existing code ...
let mut recently_sampled = self.recently_sampled.lock();
for _ in 0..batch_size {
// Try sampling with diversity enforcement (max 100 attempts)
let mut attempts = 0;
let max_attempts = 100;
let idx = loop {
let value = rng.gen::<f32>() * total_priority;
let sampled_idx = tree.sample(value)?;
// Check if this index was recently sampled
if !recently_sampled.contains(&sampled_idx) {
break sampled_idx;
}
attempts += 1;
if attempts >= max_attempts {
// Fallback: clear cooldown and use this sample
recently_sampled.clear();
break sampled_idx;
}
};
// ... process experience ...
recently_sampled.insert(idx);
}
// Increment sample counter and clear cooldown every 50 batches
let sample_count = self.sample_counter.fetch_add(1, Ordering::Relaxed);
if sample_count % 50 == 49 {
recently_sampled.clear();
}
// ... return results ...
}
3. Updated clear() Method
pub fn clear(&self) {
// ... existing clearing code ...
self.sample_counter.store(0, Ordering::Release);
// Clear diversity tracking
{
let mut recently_sampled = self.recently_sampled.lock();
recently_sampled.clear();
}
// ... reset metrics ...
}
Test Coverage
Test 1: Basic Diversity Enforcement
#[test]
fn test_batch_diversity_no_duplicates() {
// Tests that consecutive batches don't share indices
// Verifies HashSet prevents duplicate sampling
}
Test 2: Cooldown After 50 Batches
#[test]
fn test_batch_diversity_cooldown_cleared_after_50_batches() {
// Verifies cooldown is cleared after 50 batches
// Ensures long-running training doesn't accumulate all indices
}
Test 3: Fallback on Exhaustion
#[test]
fn test_batch_diversity_fallback_on_exhaustion() {
// Tests fallback when buffer size < 2*batch_size
// Ensures graceful degradation instead of infinite loop
}
Performance Characteristics
- Space: O(batch_size) for recently_sampled HashSet
- Time: O(1) expected per sample (rejection sampling), O(100) worst case
- Fallback: Clears cooldown after 100 failed attempts (prevents infinite loops)
- Memory: Cleared every 50 batches (prevents unbounded growth)
Benefits
- No duplicate experiences in consecutive batches
- Better gradient diversity - each update uses unique experiences
- Graceful fallback - handles small buffers without failing
- Automatic cleanup - cooldown reset every 50 batches
- Minimal overhead - HashSet lookups are O(1)
Edge Cases Handled
- Small buffers: Fallback clears cooldown if can't find unique samples
- Long training: Automatic cooldown reset every 50 batches
- Thread safety: Mutex protects HashSet from concurrent access
- Clear buffer: Resets both counter and cooldown set