← Alpha-v2AI AGENTS / ALPHA-V2 / METHODOLOGY / TRAINING
Methodology · § 03 / 05

Training

Alpha-v2 (model_id = 1) · by Quant Lab · Distributional SDM on liquid majors

PER. Prioritized Experience Replay with importance-sampling weights, augmented by a sign-balanced sampler that mixes UP / DOWN / FLAT clusters under fixed quotas. The clusters come from PatternMemory's KMeans over regime latents — preventing the actor from collapsing into a single sign.

Target networks. Twin target actor / critic with soft Polyak update (τ); clipped policy-smoothing noise on the target action.

Telemetry. Per-cycle segment summary: average metrics, actor/critic loss decomposition, signed and absolute force decomposition over each regulator. Sampler-health probe (actual UP/DOWN/FLAT shares vs targets — detects silent fallback to plain PER). Per-period evaluation on test windows reporting PnL, Sharpe and a composite idx.

Checkpointing. Best-by-idx + continuation checkpoints.

← PREVIOUS
Architecture
NEXT →
Regulators / Objectives
All sections of Alpha-v2
01
Overview
02
Architecture
03
Training
04
Regulators / Objectives
05
Changelog
Training — Alpha-v2 — PROGMATIC