velocitatem/PHANTOM
1
0
Fork 0
mirror of https://github.com/velocitatem/PHANTOM.git synced 2026-05-31 08:33:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

feature: telemetry logging

Browse Source
This commit is contained in:
Daniel Rosel
2026-03-10 14:23:17 +01:00
parent be03b2d4d5
commit 4c7d911043
14 changed files with 454 additions and 104 deletions
Download Patch File Download Diff File Expand all files Collapse all files

73
engine/backends/common.py
View File

@@ -19,7 +19,10 @@ def make_env(cfg: Mapping[str, Any]):
lambda_coi=float(cfg["lambda_coi"]),
robust_radius=float(cfg["robust_radius"]),
robust_points=int(cfg["robust_points"]),
robust_rollouts=int(cfg.get("robust_rollouts", 1)),
info_value=float(cfg["info_value"]),
eta_ux=float(cfg.get("eta_ux", 0.5)),
reward_profit_weight=float(cfg.get("reward_profit_weight", 1.0)),
action_levels=int(cfg["action_levels"]),
action_scale_low=float(cfg["action_scale_low"]),
action_scale_high=float(cfg["action_scale_high"]),
@@ -40,11 +43,14 @@ def _action(agent: Any, obs: Any, deterministic: bool = True):
return action
def evaluate(agent: Any, env: Any, episodes: int) -> dict[str, float]:
def _evaluate_env(agent: Any, env: Any, episodes: int) -> dict[str, float]:
rewards: list[float] = []
revenues: list[float] = []
margins: list[float] = []
coi_levels: list[float] = []
coi_leakages: list[float] = []
volatilities: list[float] = []
agent_probs: list[float] = []
for _ in range(int(episodes)):
obs, _ = env.reset()
@@ -53,6 +59,9 @@ def evaluate(agent: Any, env: Any, episodes: int) -> dict[str, float]:
ep_revenue = 0.0
ep_margin = 0.0
ep_coi = 0.0
ep_coi_leakage = 0.0
ep_volatility = 0.0
ep_agent_prob = 0.0
steps = 0
while not done:
@@ -63,6 +72,9 @@ def evaluate(agent: Any, env: Any, episodes: int) -> dict[str, float]:
ep_revenue += float(econ.get("revenue", info.get("revenue", 0.0)))
ep_margin += float(econ.get("margin", 0.0))
ep_coi += float(econ.get("coi_level", 0.0))
ep_coi_leakage += float(econ.get("coi_leakage", 0.0))
ep_volatility += float(econ.get("volatility", 0.0))
ep_agent_prob += float(econ.get("agent_prob", info.get("agent_prob", 0.0)))
steps += 1
rewards.append(ep_reward)
@@ -70,6 +82,9 @@ def evaluate(agent: Any, env: Any, episodes: int) -> dict[str, float]:
denom = max(steps, 1)
margins.append(ep_margin / denom)
coi_levels.append(ep_coi / denom)
coi_leakages.append(ep_coi_leakage / denom)
volatilities.append(ep_volatility / denom)
agent_probs.append(ep_agent_prob / denom)
return {
"eval/reward_mean": float(np.mean(rewards)) if rewards else 0.0,
@@ -78,4 +93,60 @@ def evaluate(agent: Any, env: Any, episodes: int) -> dict[str, float]:
"eval/revenue_std": float(np.std(revenues)) if revenues else 0.0,
"eval/margin_mean": float(np.mean(margins)) if margins else 0.0,
"eval/coi_level_mean": float(np.mean(coi_levels)) if coi_levels else 0.0,
"eval/coi_leakage_mean": float(np.mean(coi_leakages)) if coi_leakages else 0.0,
"eval/volatility_mean": float(np.mean(volatilities)) if volatilities else 0.0,
"eval/agent_prob_mean": float(np.mean(agent_probs)) if agent_probs else 0.0,
}
def evaluate(
agent: Any,
env: Any,
episodes: int,
cfg: Mapping[str, Any] | None = None,
) -> dict[str, float]:
metrics = _evaluate_env(agent, env, episodes)
if cfg is None or not bool(cfg.get("robust_eval_enabled", True)):
return metrics
nominal_alpha = float(cfg.get("alpha", 0.0))
eval_radius = max(float(cfg.get("robust_radius", 0.0)), 0.15)
low_alpha = float(np.clip(nominal_alpha - eval_radius, 0.0, 1.0))
high_alpha = float(np.clip(nominal_alpha + eval_radius, 0.0, 1.0))
shifted_episodes = max(1, int(np.ceil(int(episodes) / 2)))
shifted_rows = []
for tag, alpha in (
("low", low_alpha),
("nominal", nominal_alpha),
("high", high_alpha),
):
eval_cfg = dict(cfg)
eval_cfg["alpha"] = float(alpha)
shifted_env = make_env(eval_cfg)
shifted_metrics = _evaluate_env(agent, shifted_env, shifted_episodes)
shifted_env.close()
shifted_rows.append((tag, alpha, shifted_metrics))
metrics["eval/robust_alpha_low"] = low_alpha
metrics["eval/robust_alpha_high"] = high_alpha
metrics["eval/robust_reward_worst"] = float(
min(row[2]["eval/reward_mean"] for row in shifted_rows)
)
metrics["eval/robust_revenue_worst"] = float(
min(row[2]["eval/revenue_mean"] for row in shifted_rows)
)
metrics["eval/robust_coi_leakage_worst"] = float(
max(row[2]["eval/coi_leakage_mean"] for row in shifted_rows)
)
for tag, alpha, shifted_metrics in shifted_rows:
metrics[f"eval/{tag}_alpha"] = float(alpha)
metrics[f"eval/{tag}_reward_mean"] = float(shifted_metrics["eval/reward_mean"])
metrics[f"eval/{tag}_revenue_mean"] = float(
shifted_metrics["eval/revenue_mean"]
)
metrics[f"eval/{tag}_coi_leakage_mean"] = float(
shifted_metrics["eval/coi_leakage_mean"]
)
return metrics

35
engine/backends/qtable.py
View File

@@ -7,6 +7,7 @@ from typing import Any, Mapping
import numpy as np
from .common import evaluate, make_env
from ..telemetry.wandb import get_wandb_module
logger = logging.getLogger(__name__)
@@ -36,6 +37,9 @@ def train_qtable(
console_progress = bool(cfg.get("console_progress", False))
obs, _ = env.reset(seed=int(cfg["seed"]))
started_at = time.perf_counter()
wandb = get_wandb_module()
wandb_live = bool(wandb is not None and wandb.run is not None)
step_offset = max(0, int(cfg.get("wandb_step_offset", 0)))
interval_sums = {
"reward": 0.0,
@@ -75,7 +79,10 @@ def train_qtable(
"train/epsilon": float(epsilon),
"train/global_step": int(steps),
}
train_events.append(event)
if wandb_live:
wandb.log(dict(event), step=step_offset + int(steps))
else:
train_events.append(event)
if console_progress:
elapsed = max(time.perf_counter() - started_at, 1e-6)
speed = steps / elapsed
@@ -96,17 +103,19 @@ def train_qtable(
if interval_count > 0:
denom = float(interval_count)
train_events.append(
{
"train/reward_mean": interval_sums["reward"] / denom,
"train/revenue_mean": interval_sums["revenue"] / denom,
"train/agent_prob": interval_sums["agent_prob"] / denom,
"train/alpha_adv": interval_sums["alpha_adv"] / denom,
"train/coi_leakage": interval_sums["coi_leakage"] / denom,
"train/epsilon": float(epsilon),
"train/global_step": int(steps),
}
)
tail_event = {
"train/reward_mean": interval_sums["reward"] / denom,
"train/revenue_mean": interval_sums["revenue"] / denom,
"train/agent_prob": interval_sums["agent_prob"] / denom,
"train/alpha_adv": interval_sums["alpha_adv"] / denom,
"train/coi_leakage": interval_sums["coi_leakage"] / denom,
"train/epsilon": float(epsilon),
"train/global_step": int(steps),
}
if wandb_live:
wandb.log(dict(tail_event), step=step_offset + int(steps))
else:
train_events.append(tail_event)
metrics: dict[str, Any] = {
"train/reward_mean": total_reward / max(steps, 1),
@@ -114,7 +123,7 @@ def train_qtable(
"train/epsilon": float(epsilon),
"train/global_step": int(cfg["total_timesteps"]),
}
metrics.update(evaluate(agent, eval_env, int(cfg["eval_episodes"])))
metrics.update(evaluate(agent, eval_env, int(cfg["eval_episodes"]), cfg=cfg))
metrics["_train_events"] = train_events
env.close()

5
engine/backends/sb3.py
View File

@@ -144,7 +144,9 @@ def train_sb3(cfg: Mapping[str, Any]) -> tuple[object, dict[str, Any]]:
pass
metrics_callback = MetricsCallback(
log_histograms=False, log_freq=int(cfg["log_freq"])
log_histograms=False,
log_freq=int(cfg["log_freq"]),
step_offset=int(cfg.get("wandb_step_offset", 0)),
)
callbacks = [metrics_callback]
callbacks.append(
@@ -175,6 +177,7 @@ def train_sb3(cfg: Mapping[str, Any]) -> tuple[object, dict[str, Any]]:
model,
eval_env,
int(cfg["eval_episodes"]),
cfg=cfg,
)
metrics["train/global_step"] = int(model.num_timesteps)
metrics["model/path"] = str(model_path.with_suffix(".zip"))