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

fix: 10x stronger aux optimizer (LR 0.001→0.01, weight 0.01→0.1)

Browse Source 
With separate optimizer, no gradient competition — safe to increase.
Smoke test shows within-group differentiation starting (S100≠S25≠L50).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-04-08 11:13:58 +02:00
parent 6243abd971
commit af60ddf837
6 changed files with 7 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -139,7 +139,7 @@ exit_timing_weight = 0.05
ofi_reward_weight = 0.2
kelly_sizing_weight = 0.1
reward_noise_scale = 0.05
exposure_aux_weight = 0.01
exposure_aux_weight = 0.1
exposure_aux_warmup_epochs = 5
[fixed]
@@ -182,5 +182,5 @@ exit_timing_weight = 0.05
ofi_reward_weight = 0.2
kelly_sizing_weight = 0.1
reward_noise_scale = 0.05
exposure_aux_weight = 0.01
exposure_aux_weight = 0.1
exposure_aux_warmup_epochs = 5

2
config/training/dqn-localdev.toml
View File

@@ -131,5 +131,5 @@ exit_timing_weight = 0.05
ofi_reward_weight = 0.2
kelly_sizing_weight = 0.1
reward_noise_scale = 0.05
exposure_aux_weight = 0.01
exposure_aux_weight = 0.1
exposure_aux_warmup_epochs = 5

2
config/training/dqn-production.toml
View File

@@ -142,5 +142,5 @@ exit_timing_weight = 0.05
ofi_reward_weight = 0.2
kelly_sizing_weight = 0.1
reward_noise_scale = 0.05
exposure_aux_weight = 0.01
exposure_aux_weight = 0.1
exposure_aux_warmup_epochs = 5

2
config/training/dqn-smoketest.toml
View File

@@ -138,7 +138,7 @@ exit_timing_weight = 0.05
ofi_reward_weight = 0.2
kelly_sizing_weight = 0.1
reward_noise_scale = 0.05
exposure_aux_weight = 0.01
exposure_aux_weight = 0.1
exposure_aux_warmup_epochs = 5
[walk_forward]

2
crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
View File

@@ -2110,7 +2110,7 @@ impl GpuDqnTrainer {
let goff_w_b0out: usize = sizes[..10].iter().sum::<usize>();
let aux_params_ptr = self.ptrs.params_f32_ptr + (goff_w_b0out as u64) * 4;
let aux_lr = 0.001_f32; // higher LR for aux (independent from main)
let aux_lr = 0.01_f32; // 10x main LR — aux needs strong signal to break i%3
let aux_beta1 = 0.9_f32;
let aux_beta2 = 0.999_f32;
let aux_eps = 1e-8_f32;

2
crates/ml/src/trainers/dqn/config.rs
View File

@@ -1653,7 +1653,7 @@ impl DQNHyperparameters {
ofi_reward_weight: 0.2,
kelly_sizing_weight: 0.1,
reward_noise_scale: 0.05,
exposure_aux_weight: 0.01,
exposure_aux_weight: 0.1,
exposure_aux_warmup_epochs: 5,
// v8: Dense micro-reward, TD(λ), PopArt, curriculum, hindsight
micro_reward_scale: 0.01,