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

feat: position-gated episodes + 5000-bar limit — close 45x training/val gap

Browse Source 
Episode done flag: timer-based -> position-gated (trade complete = done).
V(flat)=0 is correct terminal anchor. Soft reset keeps equity on
trade completion; hard reset only on data-end or capital breach.
H100: 100 bars -> 5000 bars, gpu_n_episodes -> 1024.
ExperienceProfile gains optional gpu_n_episodes field.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-04-16 21:56:44 +02:00
parent 685746231b
commit c74a687ea8
4 changed files with 53 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
config/gpu/h100.toml
View File

@@ -7,7 +7,8 @@ hidden_dim_base = 256
replay_buffer_vram_fraction = 0.70
[experience]
gpu_timesteps_per_episode = 100
gpu_timesteps_per_episode = 5000
gpu_n_episodes = 1024
[cuda]
cuda_stack_bytes = 65536 # 64KB

7
crates/ml-core/src/gpu/profile.rs
View File

@@ -47,6 +47,9 @@ pub struct TrainingProfile {
#[derive(Debug, Clone, Deserialize)]
pub struct ExperienceProfile {
pub gpu_timesteps_per_episode: usize,
/// Number of parallel episodes (optional — training profile overrides if set).
#[serde(default)]
pub gpu_n_episodes: Option<usize>,
}
/// CUDA kernel configuration.
@@ -203,6 +206,7 @@ impl GpuProfile {
},
experience: ExperienceProfile {
gpu_timesteps_per_episode: 200,
gpu_n_episodes: None,
},
cuda: CudaProfile {
cuda_stack_bytes: 0x8000,
@@ -284,7 +288,8 @@ mod tests {
assert_eq!(profile.training.batch_size, 8192);
assert_eq!(profile.training.num_atoms, 51);
assert_eq!(profile.training.buffer_size, 500_000);
assert_eq!(profile.experience.gpu_timesteps_per_episode, 100);
assert_eq!(profile.experience.gpu_timesteps_per_episode, 5000);
assert_eq!(profile.experience.gpu_n_episodes, Some(1024));
assert_eq!(profile.cuda.cuda_stack_bytes, 65536);
}

1
crates/ml/examples/train_baseline_rl.rs
View File

@@ -424,6 +424,7 @@ fn build_dqn_hyperparams(
dataset_path: args.dataset_path.as_ref().map(|p| p.to_string_lossy().into_owned()),
replay_buffer_vram_fraction: gpu_profile.training.replay_buffer_vram_fraction,
gpu_timesteps_per_episode: gpu_profile.experience.gpu_timesteps_per_episode,
gpu_n_episodes: gpu_profile.experience.gpu_n_episodes.unwrap_or(256),
..DQNHyperparameters::default()
};

66
crates/ml/src/cuda_pipeline/experience_kernels.cu
View File

@@ -1755,7 +1755,13 @@ extern "C" __global__ void experience_env_step(
/* ---- Done detection ---- */
int next_bar = bar_idx + 1;
int done = (next_bar >= total_bars || check_capital_floor(new_portfolio_value, peak_equity)) ? 1 : 0;
/* Position-gated done: episode ends when model completes a trade cycle
* (was positioned -> now flat). V(flat)=0 is correct terminal anchor.
* Also done at end-of-data or capital floor breach (existing). */
int trade_complete = (fabsf(pre_trade_position) > 0.001f && fabsf(position) < 0.001f) ? 1 : 0;
int capital_breach = check_capital_floor(new_portfolio_value, peak_equity);
int data_end = (next_bar >= total_bars) ? 1 : 0;
int done = trade_complete | capital_breach | data_end;
/* #19 Position entropy bonus at episode end — all 4 branches */
if (done && position_histogram != NULL && position_entropy_weight > 0.0f) {
@@ -1960,28 +1966,44 @@ extern "C" __global__ void experience_env_step(
current_timesteps[i] = t + 1;
/* ---- Episode reset on done ---- */
/* When done=1, reset portfolio to fresh capital for the next episode.
* Without this, the blown portfolio persists and every subsequent step
* hits the pre-trade floor check, producing stuck episodes with fake MaxDD. */
/* Position-gated episodes: three done reasons, two reset paths.
* - data_end / capital_breach: HARD reset (full restart with fresh capital)
* - trade_complete: SOFT reset (keep equity, clear trade-tracking only)
* Soft reset lets the model accumulate gains/losses across trade cycles
* within a single data window, preventing the 45x train/val gap caused by
* resetting equity after every trade completion. */
if (done) {
float init_cap = (peak_equity); /* use peak as starting capital for next ep */
ps[0] = 0.0f; /* position = flat */
ps[1] = init_cap; /* cash */
ps[2] = init_cap; /* portfolio_value */
ps[7] = init_cap; /* peak_equity */
ps[8] = 0.0f; /* flat_counter */
ps[9] = init_cap; /* prev_equity */
ps[10] = 0.0f; /* hold_time */
ps[11] = 0.0f; /* realized_pnl */
ps[12] = 0.0f; /* entry_price */
ps[13] = 0.0f; /* trade_start_pnl */
ps[14] = 0.0f; /* win/loss/Kelly counters */
ps[15] = 0.0f;
ps[16] = 0.0f;
ps[17] = 0.0f;
ps[18] = 0.0f;
ps[19] = 0.0f;
current_timesteps[i] = 0;
if (data_end || capital_breach) {
/* Hard reset: new data window or capital wipeout — full restart */
float init_cap = (peak_equity); /* use peak as starting capital for next ep */
ps[0] = 0.0f; /* position = flat */
ps[1] = init_cap; /* cash */
ps[2] = init_cap; /* portfolio_value */
ps[7] = init_cap; /* peak_equity */
ps[8] = 0.0f; /* flat_counter */
ps[9] = init_cap; /* prev_equity */
ps[10] = 0.0f; /* hold_time */
ps[11] = 0.0f; /* realized_pnl */
ps[12] = 0.0f; /* entry_price */
ps[13] = 0.0f; /* trade_start_pnl */
ps[14] = 0.0f; /* win/loss/Kelly counters */
ps[15] = 0.0f;
ps[16] = 0.0f;
ps[17] = 0.0f;
ps[18] = 0.0f;
ps[19] = 0.0f;
current_timesteps[i] = 0;
} else {
/* Soft reset (trade_complete): keep equity, clear trade state only.
* The model starts next segment with accumulated gains/losses.
* Position is already 0 (model chose Flat). Cash, equity,
* win/loss counters, DSR state — all preserved. */
ps[10] = 0.0f; /* hold_time — new trade starts fresh */
ps[12] = 0.0f; /* entry_price — no active trade */
ps[13] = (ps[11]); /* trade_start_pnl = current realized_pnl */
ps[20] = 0.0f; /* intra_trade_max_dd — reset for next trade */
current_timesteps[i] = 0;
}
}
}