119 lines
4.1 KiB
Python
119 lines
4.1 KiB
Python
"""Stable, reward-weight-independent evaluation scoring."""
|
||
|
||
from __future__ import annotations
|
||
|
||
import math
|
||
from collections.abc import Mapping
|
||
from typing import Any
|
||
|
||
DEFAULT_OBJECTIVE_WEIGHTS = {
|
||
"velocity_tracking": 0.35,
|
||
"action_smoothness": 0.20,
|
||
"posture_stability": 0.15,
|
||
"fall_avoidance": 0.15,
|
||
"foot_slip": 0.10,
|
||
"energy": 0.05,
|
||
}
|
||
|
||
REQUIRED_METRICS = {
|
||
"linear_velocity_rmse",
|
||
"angular_velocity_rmse",
|
||
"mean_action_acc",
|
||
"orientation_error",
|
||
"fall_rate",
|
||
"slip_velocity",
|
||
"mechanical_power",
|
||
}
|
||
|
||
PHYSICAL_FLOORS = {
|
||
"linear_velocity_rmse": 0.10,
|
||
"angular_velocity_rmse": 0.10,
|
||
"mean_action_acc": 0.01,
|
||
"orientation_error": 0.05,
|
||
"fall_rate": 0.02,
|
||
"slip_velocity": 0.05,
|
||
"mechanical_power": 10.0,
|
||
}
|
||
|
||
|
||
class EvaluationError(ValueError):
|
||
pass
|
||
|
||
|
||
def validate_objective_weights(value: Any) -> dict[str, float]:
|
||
if not isinstance(value, Mapping) or set(value) != set(DEFAULT_OBJECTIVE_WEIGHTS):
|
||
raise EvaluationError("objectiveWeights 必须完整包含六个目标")
|
||
result: dict[str, float] = {}
|
||
for key in DEFAULT_OBJECTIVE_WEIGHTS:
|
||
raw = value[key]
|
||
if isinstance(raw, bool) or not isinstance(raw, (int, float)):
|
||
raise EvaluationError(f"objectiveWeights.{key} 必须是数值")
|
||
number = float(raw)
|
||
if not math.isfinite(number) or number < 0.0 or number > 1.0:
|
||
raise EvaluationError(f"objectiveWeights.{key} 必须在 0–1 之间")
|
||
result[key] = number
|
||
if not math.isclose(sum(result.values()), 1.0, abs_tol=1.0e-6):
|
||
raise EvaluationError("objectiveWeights 总和必须为 1")
|
||
return result
|
||
|
||
|
||
def validate_metrics(value: Any) -> dict[str, float]:
|
||
if not isinstance(value, Mapping):
|
||
raise EvaluationError("metrics 必须是对象")
|
||
missing = REQUIRED_METRICS - set(value)
|
||
if missing:
|
||
raise EvaluationError(f"metrics 缺少：{', '.join(sorted(missing))}")
|
||
result: dict[str, float] = {}
|
||
for key in REQUIRED_METRICS:
|
||
raw = value[key]
|
||
if isinstance(raw, bool) or not isinstance(raw, (int, float)):
|
||
raise EvaluationError(f"metrics.{key} 必须是数值")
|
||
number = float(raw)
|
||
if not math.isfinite(number) or number < 0.0:
|
||
raise EvaluationError(f"metrics.{key} 必须是非负有限数值")
|
||
result[key] = number
|
||
if result["fall_rate"] > 1.0:
|
||
raise EvaluationError("metrics.fall_rate 必须在 0–1 之间")
|
||
return result
|
||
|
||
|
||
def _improvement(baseline: Mapping[str, float], current: Mapping[str, float], key: str) -> float:
|
||
scale = max(abs(baseline[key]), PHYSICAL_FLOORS[key])
|
||
return max(-1.0, min(1.0, (baseline[key] - current[key]) / scale))
|
||
|
||
|
||
def score_evaluation(
|
||
baseline_value: Any,
|
||
current_value: Any,
|
||
objective_weights: Any = DEFAULT_OBJECTIVE_WEIGHTS,
|
||
) -> dict[str, Any]:
|
||
baseline = validate_metrics(baseline_value)
|
||
current = validate_metrics(current_value)
|
||
weights = validate_objective_weights(objective_weights)
|
||
components = {
|
||
"velocity_tracking": 0.8 * _improvement(baseline, current, "linear_velocity_rmse")
|
||
+ 0.2 * _improvement(baseline, current, "angular_velocity_rmse"),
|
||
"action_smoothness": _improvement(baseline, current, "mean_action_acc"),
|
||
"posture_stability": _improvement(baseline, current, "orientation_error"),
|
||
"fall_avoidance": _improvement(baseline, current, "fall_rate"),
|
||
"foot_slip": _improvement(baseline, current, "slip_velocity"),
|
||
"energy": _improvement(baseline, current, "mechanical_power"),
|
||
}
|
||
tracking_limit = max(
|
||
baseline["linear_velocity_rmse"] * 1.05, baseline["linear_velocity_rmse"] + 1.0e-6
|
||
)
|
||
eligible = (
|
||
current["fall_rate"] <= baseline["fall_rate"] + 0.02
|
||
and current["linear_velocity_rmse"] <= tracking_limit
|
||
)
|
||
total = sum(weights[key] * components[key] for key in weights)
|
||
if not eligible:
|
||
total = min(total, -1.0)
|
||
return {
|
||
"eligible": eligible,
|
||
"score": total,
|
||
"components": components,
|
||
"metrics": current,
|
||
"baselineMetrics": baseline,
|
||
}
|