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 e97c30b50a
commit 41eaa6522c
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
crates/ml/src/hyperopt/adapters/dqn.rs
View File

@@ -3719,7 +3719,7 @@ mod tests {
assert_eq!(bounds[11], (1.0, 1.0)); // use_branching (FIXED: always true — GPU pipeline requires branching)
assert_eq!(bounds[13], (128.0, 512.0)); // dueling_hidden_dim
assert_eq!(bounds[14], (3.0, 5.0)); // n_steps (raised min: Rainbow standard)
assert_eq!(bounds[15], (11.0, 101.0)); // num_atoms (11=RTX3050, 101=max)
assert_eq!(bounds[15], (11.0, 51.0)); // num_atoms (11=RTX3050, 51=H100 default)
// Kelly risk parameters
assert_eq!(bounds[17], (0.25, 0.75)); // kelly_fractional
@@ -3727,7 +3727,7 @@ mod tests {
assert_eq!(bounds[19], (10.0, 30.0)); // volatility_window
// C2: curiosity_weight fixed to 0.0, noisy_epsilon_floor fixed to 0.10, not in search space
assert_eq!(bounds[21], (128.0, 512.0)); // hidden_dim_base (128-512, backtest is bottleneck not training)
assert_eq!(bounds[21], (128.0, 256.0)); // hidden_dim_base (capped at production default for hyperopt speed)
assert_eq!(bounds[22], (0.0, 1.0)); // cql_alpha (full offline-RL range)
// C5: branch_hidden_dim