← Alpha-v2.4AI AGENTS / ALPHA-V2.4 / METHODOLOGY / REGULATORS / OBJECTIVES
Methodology · § 04 / 05

Regulators / Objectives

Alpha-v2.4 (model_id = 4) · by Quant Lab · N-agnostic refactor — branch / V_global separation

Branch regulators (per-symbol).

action_l2_coeff. Anti-saturation: relu(|a| − threshold)² × coeff.

portfolio_budget_quad_coeff. Quadratic penalty for gross exposure exceeding the limit.

inaction_penalty_coeff. Anti-zero-policy.

action_diversity_coeff. Anti-symbol-collapse via std(actions, dim=symbols).

direction_inbalance_coeff. Anti-mono-direction — penalty when |batch_mean(actions)| exceeds threshold.

q_crossing_coeff. Penalty for monotonicity violations of the per-sym quantile distribution (numerical stability).

New v2.4 knobs.

actor_mini_mixer_layers / heads / ffn_mult — capacity of the actor cross-symbol mixer.

v_global_layers / heads / pool_query_dim — capacity of the V_global Transformer observer.

v_global_loss_coeff — weight of the V_global TD term in the total critic objective.

Deprecated / removed.

encoder_grad_source — superseded by the v2.3 separate-encoder split (always td_only on its own path).

actor_head_arch — only the linear variant remains; ResidualSkipHead has been removed.

regime_adapter_* — actor_adapter was removed entirely.

← PREVIOUS
Training
NEXT →
Changelog
All sections of Alpha-v2.4
01
Overview
02
Architecture
03
Training
04
Regulators / Objectives
05
Changelog
Regulators / Objectives — Alpha-v2.4 — PROGMATIC