File size: 12,521 Bytes
6f90f5c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 | """Feedback loop for eval agent — computes effectiveness of auxiliary metrics."""
from __future__ import annotations
import json
import logging
import math
from pathlib import Path
from typing import Dict, List, Optional, Tuple
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Statistics helpers (no scipy dependency)
# ---------------------------------------------------------------------------
def _rank(values: List[float]) -> List[float]:
"""Assign ranks to values (average rank for ties)."""
n = len(values)
indexed = sorted(range(n), key=lambda i: values[i])
ranks = [0.0] * n
i = 0
while i < n:
j = i
while j < n - 1 and values[indexed[j + 1]] == values[indexed[j]]:
j += 1
avg_rank = (i + j) / 2.0 + 1 # 1-based
for k in range(i, j + 1):
ranks[indexed[k]] = avg_rank
i = j + 1
return ranks
def _spearman_correlation(x: List[float], y: List[float]) -> float:
"""Spearman rank correlation without scipy."""
n = len(x)
if n < 3:
return 0.0
rx = _rank(x)
ry = _rank(y)
d_sq = sum((a - b) ** 2 for a, b in zip(rx, ry))
denom = n * (n * n - 1)
if denom == 0:
return 0.0
return 1.0 - (6.0 * d_sq) / denom
def _variance(values: List[float]) -> float:
"""Population variance."""
n = len(values)
if n < 2:
return 0.0
mean = sum(values) / n
return sum((v - mean) ** 2 for v in values) / n
def _linear_trend(values: List[float]) -> float:
"""Slope of linear regression (values indexed 0..n-1)."""
n = len(values)
if n < 2:
return 0.0
x_mean = (n - 1) / 2.0
y_mean = sum(values) / n
num = sum((i - x_mean) * (v - y_mean) for i, v in enumerate(values))
den = sum((i - x_mean) ** 2 for i in range(n))
if den == 0:
return 0.0
return num / den
# ---------------------------------------------------------------------------
# Data loading
# ---------------------------------------------------------------------------
def _load_generation_metrics(
results_dir: Path, gen_start: int, gen_end: int
) -> List[Tuple[int, Dict]]:
"""Load metrics.json for a range of generations. Returns [(gen, data), ...]."""
records = []
for gen in range(gen_start, gen_end + 1):
metrics_path = results_dir / f"gen_{gen}" / "results" / "metrics.json"
if not metrics_path.exists():
continue
try:
with open(metrics_path) as f:
data = json.load(f)
records.append((gen, data))
except Exception:
continue
return records
def _extract_aux_metric_names(records: List[Tuple[int, Dict]]) -> List[str]:
"""Collect all aux_ metric names from public metrics across records.
Filters out framework metrics (aux_aux_metric_*) that are always present."""
FRAMEWORK_PREFIXES = ("aux_aux_metric_eval_success", "aux_aux_metric_error_code",
"aux_aux_metric_error_message_length", "aux_aux_metric_error_detail_length",
"aux_aux_metric_non_numeric_dropped_count")
names: set[str] = set()
for _, data in records:
public = data.get("public", {})
for key in public:
if key.startswith("aux_") and key not in FRAMEWORK_PREFIXES:
names.add(key)
return sorted(names)
# ---------------------------------------------------------------------------
# Feedback computation
# ---------------------------------------------------------------------------
def _classify_correlation(rho: float) -> Tuple[str, str]:
"""Classify correlation strength. Returns (label, emoji)."""
abs_rho = abs(rho)
if abs_rho >= 0.6:
if rho > 0:
return "strong positive", "USEFUL"
return "strong negative", "MISLEADING"
if abs_rho >= 0.3:
if rho > 0:
return "moderate positive", "USEFUL"
return "moderate negative", "REVIEW"
return "weak/none", "LOW SIGNAL"
def compute_metric_feedback(
results_dir: Path,
current_gen: int,
last_agent_gen: int,
pre_window: int = 10,
) -> str:
"""
Compute feedback on auxiliary metrics effectiveness since last agent intervention.
Args:
results_dir: Experiment root directory.
current_gen: Current generation number.
last_agent_gen: Generation when agent last ran.
pre_window: Number of generations before agent intervention to compare.
Returns:
Formatted markdown string for inclusion in agent task message.
Empty string if insufficient data.
"""
if last_agent_gen < 0 or current_gen <= last_agent_gen:
return ""
results_dir = Path(results_dir)
# Load post-intervention data (last_agent_gen+1 .. current_gen)
post_records = _load_generation_metrics(results_dir, last_agent_gen + 1, current_gen)
if len(post_records) < 2:
return "" # Not enough data for meaningful feedback
# Load pre-intervention data for trajectory comparison
pre_start = max(0, last_agent_gen - pre_window)
pre_records = _load_generation_metrics(results_dir, pre_start, last_agent_gen)
# --- Section 1: Score trajectory comparison ---
trajectory_lines = _build_trajectory_section(pre_records, post_records, last_agent_gen)
# --- Section 2: Per-metric effectiveness ---
aux_names = _extract_aux_metric_names(post_records)
if not aux_names:
# No auxiliary metrics to evaluate
if trajectory_lines:
return (
f"\n\n📋 Feedback on Previous Actions (gen {last_agent_gen} → gen {current_gen}):\n\n"
+ "\n".join(trajectory_lines)
+ "\n\nNo auxiliary metrics found in recent generations.\n"
)
return ""
metric_lines = _build_metric_effectiveness_section(post_records, aux_names)
# --- Section 3: Actionable recommendation ---
recommendation = _build_recommendation(post_records, aux_names)
# Assemble
parts = [
f"\n\n📋 Feedback on Your Previous Actions (gen {last_agent_gen} → gen {current_gen}):",
"",
]
if trajectory_lines:
parts.extend(trajectory_lines)
parts.append("")
parts.extend(metric_lines)
if recommendation:
parts.append("")
parts.append(recommendation)
return "\n".join(parts)
def _build_trajectory_section(
pre_records: List[Tuple[int, Dict]],
post_records: List[Tuple[int, Dict]],
last_agent_gen: int,
) -> List[str]:
"""Build score trajectory comparison lines."""
lines: List[str] = ["Score Trajectory:"]
pre_scores = [d.get("combined_score", 0.0) for _, d in pre_records if isinstance(d.get("combined_score"), (int, float))]
post_scores = [d.get("combined_score", 0.0) for _, d in post_records if isinstance(d.get("combined_score"), (int, float))]
if pre_scores:
pre_avg = sum(pre_scores) / len(pre_scores)
pre_trend = _linear_trend(pre_scores)
lines.append(f"- Before intervention ({len(pre_scores)} gens): avg {pre_avg:.4f}, trend {pre_trend:+.4f}/gen")
if post_scores:
post_avg = sum(post_scores) / len(post_scores)
post_trend = _linear_trend(post_scores)
lines.append(f"- After intervention ({len(post_scores)} gens): avg {post_avg:.4f}, trend {post_trend:+.4f}/gen")
if pre_scores and post_scores:
delta = post_avg - pre_avg
lines.append(f"- Net impact: avg score {'improved' if delta > 0 else 'declined'} by {delta:+.4f}")
# Score decline warning
if pre_avg > 0 and delta < 0 and abs(delta) / pre_avg > 0.05:
lines.append("")
lines.append("⚠️ WARNING: Score has DECLINED significantly since agent intervention.")
lines.append("Consider simplifying or clearing auxiliary metrics to eliminate potential interference.")
lines.append("Focus diagnostic_report.md on algorithmic advice rather than adding new metrics.")
return lines if len(lines) > 1 else []
def _build_metric_effectiveness_section(
post_records: List[Tuple[int, Dict]],
aux_names: List[str],
) -> List[str]:
"""Build per-metric effectiveness table."""
scores = []
for _, data in post_records:
cs = data.get("combined_score")
if isinstance(cs, (int, float)) and math.isfinite(cs):
scores.append(cs)
else:
scores.append(None)
lines = [
"Metric Effectiveness Report:",
"| Metric | Corr w/ Score | Variance | Coverage | Signal |",
"|--------|--------------|----------|----------|--------|",
]
for name in aux_names:
metric_vals: List[Optional[float]] = []
for _, data in post_records:
v = data.get("public", {}).get(name)
if isinstance(v, (int, float)) and math.isfinite(v):
metric_vals.append(float(v))
else:
metric_vals.append(None)
# Paired non-None values for correlation
paired_m = []
paired_s = []
for mv, sv in zip(metric_vals, scores):
if mv is not None and sv is not None:
paired_m.append(mv)
paired_s.append(sv)
n_total = len(metric_vals)
n_present = sum(1 for v in metric_vals if v is not None)
coverage = n_present / n_total if n_total > 0 else 0.0
if len(paired_m) >= 3:
rho = _spearman_correlation(paired_m, paired_s)
var = _variance(paired_m)
label, signal = _classify_correlation(rho)
# Override signal if variance is near zero
if var < 1e-9:
signal = "LOW SIGNAL"
lines.append(
f"| {name} | {rho:+.2f} ({label}) | {var:.4f} | {coverage:.0%} | {signal} |"
)
else:
lines.append(
f"| {name} | N/A (< 3 pts) | N/A | {coverage:.0%} | INSUFFICIENT |"
)
return lines
def _build_recommendation(
post_records: List[Tuple[int, Dict]],
aux_names: List[str],
) -> str:
"""Build actionable recommendation based on metric analysis."""
scores = [
data.get("combined_score", 0.0)
for _, data in post_records
if isinstance(data.get("combined_score"), (int, float))
]
to_remove: List[str] = []
to_review: List[str] = []
for name in aux_names:
paired_m = []
paired_s = []
for _, data in post_records:
v = data.get("public", {}).get(name)
cs = data.get("combined_score")
if (
isinstance(v, (int, float)) and math.isfinite(v)
and isinstance(cs, (int, float)) and math.isfinite(cs)
):
paired_m.append(float(v))
paired_s.append(float(cs))
if len(paired_m) < 3:
continue
rho = _spearman_correlation(paired_m, paired_s)
var = _variance(paired_m)
if var < 1e-9:
# Distinguish: constant because code broken vs constant because metric is bad
all_scores_zero = all(s <= 0.01 for s in paired_s)
if all_scores_zero:
# Don't recommend removal — metric is constant because code never worked
pass # skip this metric, it may become useful when code starts working
else:
to_remove.append(f"{name} (constant value, no signal)")
elif rho < -0.3:
to_review.append(f"{name} (negative correlation {rho:+.2f}, may mislead optimizer)")
elif abs(rho) < 0.1:
to_remove.append(f"{name} (near-zero correlation {rho:+.2f})")
parts: List[str] = []
if to_remove:
parts.append("Consider REMOVING: " + "; ".join(to_remove) + ".")
if to_review:
parts.append("Consider REVIEWING: " + "; ".join(to_review) + ".")
if not parts:
parts.append("All current metrics show reasonable signal. Focus on diagnostic_report.md rather than adding new metrics.")
# Churn warning: if most metrics are flagged for removal, the agent is likely churning
if len(to_remove) >= 3:
parts.append(
"⚠️ CHURN WARNING: Most metrics have low signal. "
"Stop modifying metrics and focus on writing a useful diagnostic_report.md instead."
)
return "Recommendation: " + " ".join(parts)
|