45879f3901
Agent 3 — set_count_bonuses HOT path (3-6 HtoD/call → 0), gpu_dqn_trainer constructor block (11 COLD sites: weight_decay_mask, branch metadata, gamma, q_quantile, spectral_norm, stochastic_depth, VSN groups, mamba2 init), upload_params/upload_target_params (2 WARM), HER + curriculum episode_starts (3 WARM sites in fused_training and training_loop). Strategy: COLD/WARM sites use mapped-pinned staging + DtoD into existing CudaSlice destinations to avoid disturbing 100+ downstream consumers; HOT set_count_bonuses uses persistent MappedF32Buffer fields (no DtoD per call). 5 commits, 23 sites total. Adds MappedU64Buffer for spectral-norm host_desc[78]. Per feedback_no_htod_htoh_only_mapped_pinned.md — third of 3 parallel batches. # Conflicts: # crates/ml/src/cuda_pipeline/mapped_pinned.rs
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;