b5d19c1004
Plan 3 Task 3.
ISV tail-append:
- [71] TRADE_ATTEMPT_RATE_EMA — Flat→Positioned transition rate EMA
(GPU-written by trade_attempt_rate_ema_update, adaptive α)
- [72] TRADE_TARGET_RATE — reference rate, CPU-frozen at epoch 5
- Fingerprint shifted [69,70] → [73,74]; ISV_TOTAL_DIM 71 → 75
Producer kernel (trade_rate_ema_kernel.cu):
- Single-block reduction of flat_to_pos_per_sample [N*L] (no atomicAdd)
- Adaptive EMA: α = α_base × (1 + 0.5 × |clamp(sharpe, -2, 2)|)
- α_base = 0.05 (matches reward_component_ema convention)
- Launched from training_loop alongside reward_component_ema
Consumer (experience_env_step):
- Flat→Positioned site: novelty = max(0, 1 - attempt/target)
- bonus = conviction_core × vol_proxy × novelty
- reward += shaping_scale × bonus; rc[5] captures bonus for ISV[68]
- Explicit freeze gate: target_raw > 1e-6f, so bonus is structurally
inert pre-freeze (prevents spurious novelty=1.0 on epoch-1 when
attempt_rate is still 0)
Epoch-5 freeze (training_loop):
- measured = ISV[TRADE_ATTEMPT_RATE_EMA]; floor at 0.001
- Prevents novelty from sticking at 1.0 post-freeze
StateResetRegistry: both slots registered as FoldReset with per-slot
reset dispatch arms in training_loop's fold-boundary path.
Smoke: multi_fold_convergence passes (fold 2 best Sharpe 87.55 —
slightly above Task 2 baseline 85.6, within noise; bonus inert
in 5-epoch smoke so training matches pre-B.2 baseline as designed).
cargo check clean at 11 warnings baseline.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ml
10-model ML ensemble for the Foxhunt HFT system, built on Candle v0.9.1.
Models
- DQN (Rainbow) — deep Q-network with prioritized replay, dueling heads, noisy nets
- PPO — proximal policy optimization with GAE, LSTM policies, clip-higher
- TFT — temporal fusion transformer for multi-horizon forecasting
- Mamba2 — state space model for sequence prediction
- Liquid Networks — biologically inspired networks for non-stationary data
- TLOB — transformer-based limit order book analysis
- KAN — Kolmogorov-Arnold networks
- xLSTM — extended LSTM architecture
- TGGN — temporal graph neural network
- Diffusion — diffusion-based generative model
Key Modules
ensemble— model ensemble coordination and confidence aggregationhyperopt— PSO-based hyperparameter optimization with per-model adapterstrainers— unified training loops (DQN, PPO, supervised)inference—InferenceAdaptertrait for predictioncheckpoint— model checkpointing and restorationevaluation— walk-forward evaluation pipeline
Usage
use ml::dqn::DQN;
use ml::ppo::PpoTrainer;