gitadmin/foxhunt
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

perf: cap hyperopt hidden_dim=256, num_atoms=51 — 4x faster trials

Browse Source 
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) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-03-22 10:00:26 +01:00
parent 11855e958a
commit e97c30b50a
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
config/training/dqn-hyperopt.toml
View File

@@ -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]