From e97c30b50a2aa3cccd7b45bb5e9672b478773fd5 Mon Sep 17 00:00:00 2001 From: jgrusewski Date: Sun, 22 Mar 2026 10:00:26 +0100 Subject: [PATCH] =?UTF-8?q?perf:=20cap=20hyperopt=20hidden=5Fdim=3D256,=20?= =?UTF-8?q?num=5Fatoms=3D51=20=E2=80=94=204x=20faster=20trials?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Production defaults are sufficient for hyperopt exploration. Larger networks can be tested in a separate phase with the best hyperparams found. Co-Authored-By: Claude Opus 4.6 (1M context) --- config/training/dqn-hyperopt.toml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/config/training/dqn-hyperopt.toml b/config/training/dqn-hyperopt.toml index 41e140eb6..0bcecc5ca 100644 --- a/config/training/dqn-hyperopt.toml +++ b/config/training/dqn-hyperopt.toml @@ -20,7 +20,7 @@ v_range = [10.0, 50.0] # symmetric: v_min=-v_range, v_max=+v_ran noisy_sigma_init = [0.1, 1.0] # log scale in adapter dueling_hidden_dim = [128, 512] n_steps = [3, 5] -num_atoms = [11, 101] # 11=RTX3050, 51=H100 default, 101=max +num_atoms = [11, 51] # 11=RTX3050, 51=H100 default (101 is 2x slower for marginal benefit) # Weight decay weight_decay = [0.0001, 0.01] # log scale in adapter @@ -36,7 +36,7 @@ volatility_window = [10, 30] tau = [0.005, 0.01] # log scale in adapter # Network sizing -hidden_dim_base = [128, 512] # capped: backtest evaluator is O(bars × dim²) +hidden_dim_base = [128, 256] # capped at production default: 512 is 4x slower for marginal benefit # CQL regularization cql_alpha = [0.0, 1.0]