Spaces:
Sleeping
Sleeping
File size: 18,336 Bytes
b14c6e3 1341fa9 b14c6e3 d6fcbb0 b14c6e3 d6fcbb0 b14c6e3 c18a9d1 b14c6e3 c18a9d1 b14c6e3 c18a9d1 1341fa9 b14c6e3 7cd2458 b14c6e3 7cd2458 b14c6e3 7ee6ca2 7cd2458 b14c6e3 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 | """
Task 2: Resource-Constrained Triage (Medium)
=============================================
Objective
---------
Maximise incident resolution quality under a hard per-step investigation
budget (K = 3 INVESTIGATE actions per step). The agent must learn to
prioritise which alerts deserve expensive investigations versus which can be
cheaply ignored or escalated.
Difficulty contract
-------------------
- max_investigations = 3 per step (enforced by env.py)
- Alerts are still mostly independent (20 % chance of correlated chain)
- False-positive rate elevated vs easy task
- Agent must balance breadth (handle many alerts) vs depth (investigate few)
Grading formula (all inputs are deterministic given the same seed)
-------------------------------------------------------------------
resolved_score = Ξ£ severity_weight(alert) Β· resolution_quality(action)
for all handled alerts
fp_penalty = 0.30 Β· (unnecessary_investigations / total_investigations)
β fraction of budget spent on false positives
critical_miss_penalty = 0.20 Β· min(critical_missed / max(critical_total, 1), 1.0)
raw = resolved_score / max_possible_score β [0, 1]
score = max(0.0, raw β fp_penalty β critical_miss_penalty) β [0.0, 1.0]
Resolution quality per action
------------------------------
INVESTIGATE on critical/high (β₯ 0.75) β 1.00 Γ severity
INVESTIGATE on medium (0.50β0.74) β 0.80 Γ severity
INVESTIGATE on false-pos (< 0.30) β 0.00 + marks unnecessary
ESCALATE on critical/high(β₯ 0.70) β 0.85 Γ severity
ESCALATE on medium β 0.60 Γ severity
ESCALATE on false-pos β 0.00 + marks unnecessary
IGNORE on false-pos (< 0.30) β 0.15 Γ (1 β severity) (filtering bonus)
IGNORE on medium β 0.00
IGNORE on critical β 0.00 + critical_missed++
DELAY on any β 0.00 (deferral, no score now)
Success threshold: β₯ 0.55
Integration with env.py
-----------------------
Consume info["processed_alerts"] inside the step loop:
obs, reward, done, info = env.step(action)
for alert_data in info.get("processed_alerts", []):
grader.process_step(alert_data, info)
"""
from __future__ import annotations
from typing import Any, Dict, List
import numpy as np
from adaptive_alert_triage.models import Action, Alert, Reward
# ---------------------------------------------------------------------------
# Constants
# ---------------------------------------------------------------------------
_CRITICAL_THRESHOLD: float = 0.75
_HIGH_THRESHOLD: float = 0.70
_MEDIUM_LOWER: float = 0.50
_FALSE_POSITIVE_THRESHOLD: float = 0.30
# Grading weights
_FP_PENALTY_WEIGHT: float = 0.30
_CRITICAL_MISS_PENALTY_WEIGHT: float = 0.20
# Filtering-bonus cap so ignoring FPs never inflates score above 1.0
_FP_BONUS_CAP_PER_ALERT: float = 0.15
SUCCESS_THRESHOLD: float = 0.549
# ---------------------------------------------------------------------------
# Grader
# ---------------------------------------------------------------------------
class MediumTaskGrader:
"""
Grader for Task 2: Resource-Constrained Triage.
Lifecycle (one episode)
-----------------------
1. Instantiate once per episode.
2. After every env.step(action), iterate info["processed_alerts"] and
call process_step(alert_data, info) for each entry.
3. At episode end call get_episode_score() β float in [0.0, 1.0].
4. Optionally call get_metrics() for a full breakdown.
5. Call reset() to reuse for a new episode.
The score is deterministic: same seed + same policy β same score.
"""
def __init__(self, max_investigations_per_step: int = 3) -> None:
self._K = max_investigations_per_step
# Accumulators
self._resolved_score: float = 0.0 # weighted resolution quality
self._max_possible_score: float = 0.0 # theoretical max if all handled optimally
self._total_investigations: int = 0
self._unnecessary_invest: int = 0 # INVESTIGATE on FP or low severity
self._critical_total: int = 0
self._critical_missed: int = 0
self._total_actions: int = 0
self._action_history: List[Dict[str, Any]] = []
# ------------------------------------------------------------------
# Primary interface
# ------------------------------------------------------------------
def process_step(
self,
alert_data: Dict[str, Any],
info: Dict[str, Any], # noqa: ARG002
) -> float:
"""
Evaluate one action using ground-truth data from env.step().
Args:
alert_data: One entry from info["processed_alerts"].
info: Full info dict (unused here, kept for API symmetry).
Returns:
Raw score contribution for this action (not normalised).
"""
self._total_actions += 1
true_sev: float = float(alert_data.get("true_severity", 0.0))
action_type: str = str(alert_data.get("action_taken", ""))
is_fp: bool = bool(alert_data.get("is_false_positive",
true_sev < _FALSE_POSITIVE_THRESHOLD))
# The theoretical max contribution for this alert (investigating optimally)
optimal = self._optimal_contribution(true_sev)
self._max_possible_score += optimal
if true_sev >= _CRITICAL_THRESHOLD:
self._critical_total += 1
contribution = self._contribution(action_type, true_sev, is_fp)
self._resolved_score += contribution
if action_type == "INVESTIGATE":
self._total_investigations += 1
if is_fp or true_sev < _FALSE_POSITIVE_THRESHOLD:
self._unnecessary_invest += 1
if action_type == "IGNORE" and true_sev >= _CRITICAL_THRESHOLD:
self._critical_missed += 1
self._action_history.append({
"alert_id": alert_data.get("alert_id", ""),
"action": action_type,
"true_severity": true_sev,
"is_fp": is_fp,
"contribution": contribution,
"optimal": optimal,
})
return contribution
# ------------------------------------------------------------------
# Legacy API
# ------------------------------------------------------------------
def grade_action(self, action: Action, alert: Alert, reward: Reward) -> float:
"""Grade a single action-alert pair (legacy / unit-test API)."""
alert_data = {
"alert_id": alert.id,
"true_severity": alert.true_severity,
"visible_severity": alert.visible_severity,
"confidence": alert.confidence,
"alert_type": alert.alert_type,
"age": alert.age,
"action_taken": action.action_type,
"is_false_positive": alert.true_severity < _FALSE_POSITIVE_THRESHOLD,
}
return self.process_step(alert_data, {})
# ------------------------------------------------------------------
# Scoring
# ------------------------------------------------------------------
def get_episode_score(self) -> float:
"""
Return final normalised score strictly in (0, 1) β never 0.0 or 1.0.
"""
if self._max_possible_score <= 0.0:
return 0.5
raw = self._resolved_score / self._max_possible_score
if self._total_investigations > 0:
fp_rate = self._unnecessary_invest / self._total_investigations
else:
fp_rate = 0.0
fp_penalty = _FP_PENALTY_WEIGHT * fp_rate
if self._critical_total > 0:
miss_rate = min(self._critical_missed / self._critical_total, 1.0)
else:
miss_rate = 0.0
miss_penalty = _CRITICAL_MISS_PENALTY_WEIGHT * miss_rate
# Penalised score is effectively between -0.50 and 1.00
penalised = raw - fp_penalty - miss_penalty
# Math map: penalised * 0.6 is [-0.3, 0.6]
# + 0.35 yields [0.05, 0.95] which guarantees (0, 1) bounds without clipping.
mapped = (penalised * 0.6) + 0.35
return float(round(mapped, 4))
def passed(self) -> bool:
"""Return True if the agent meets the medium-task success threshold."""
return self.get_episode_score() >= SUCCESS_THRESHOLD
def calculate_resource_efficiency(self) -> float:
"""
Fraction of INVESTIGATE + ESCALATE actions that were productive.
Productive = action on an alert with true_severity β₯ 0.50.
Returns 1.0 when no costly actions were taken (or 1.0 for perfect efficiency).
"""
costly = [h for h in self._action_history
if h["action"] in ("INVESTIGATE", "ESCALATE")]
if not costly:
return 1.0
productive = sum(1 for h in costly if h["true_severity"] >= _MEDIUM_LOWER)
raw = productive / len(costly)
return raw
# ------------------------------------------------------------------
# Metrics
# ------------------------------------------------------------------
def get_metrics(self) -> Dict[str, Any]:
"""Return a full breakdown of episode performance."""
score = self.get_episode_score()
eff = self.calculate_resource_efficiency()
breakdown: Dict[str, int] = {
"INVESTIGATE": 0, "IGNORE": 0, "ESCALATE": 0, "DELAY": 0,
}
for h in self._action_history:
breakdown[h["action"]] = breakdown.get(h["action"], 0) + 1
return {
"overall_score": score,
"passed": self.passed(),
"success_threshold": SUCCESS_THRESHOLD,
"resolved_score": self._resolved_score,
"max_possible_score": self._max_possible_score,
"normalised_resolved": (self._resolved_score / self._max_possible_score
if self._max_possible_score > 0 else 0.0),
"resource_efficiency": eff,
"total_investigations": self._total_investigations,
"unnecessary_invest": self._unnecessary_invest,
"critical_total": self._critical_total,
"critical_missed": self._critical_missed,
"total_actions": self._total_actions,
"action_breakdown": breakdown,
}
# ------------------------------------------------------------------
# Housekeeping
# ------------------------------------------------------------------
def reset(self) -> None:
"""Reset all state for a new episode."""
self._resolved_score = 0.0
self._max_possible_score = 0.0
self._total_investigations = 0
self._unnecessary_invest = 0
self._critical_total = 0
self._critical_missed = 0
self._total_actions = 0
self._action_history = []
def __repr__(self) -> str:
score = self.get_episode_score()
eff = self.calculate_resource_efficiency()
return (
f"MediumTaskGrader(score={score:.3f}, "
f"efficiency={eff:.3f}, "
f"investigations={self._total_investigations}, "
f"passed={self.passed()})"
)
# ------------------------------------------------------------------
# Private helpers
# ------------------------------------------------------------------
@staticmethod
def _optimal_contribution(true_sev: float) -> float:
"""What's the best possible contribution for this alert?"""
if true_sev >= _CRITICAL_THRESHOLD:
return 1.00 * true_sev
if true_sev >= _MEDIUM_LOWER:
return 0.80 * true_sev
if true_sev < _FALSE_POSITIVE_THRESHOLD:
return _FP_BONUS_CAP_PER_ALERT * (1.0 - true_sev)
# Low-medium: best action is still INVESTIGATE
return 0.80 * true_sev
@staticmethod
def _contribution(action_type: str, true_sev: float, is_fp: bool) -> float:
"""
Deterministic contribution for one action.
Returns a non-normalised float; caller accumulates into
_resolved_score and later normalises by _max_possible_score.
"""
if action_type == "INVESTIGATE":
if is_fp or true_sev < _FALSE_POSITIVE_THRESHOLD:
return 0.0 # budget wasted; penalty applied separately
if true_sev >= _CRITICAL_THRESHOLD:
return 1.00 * true_sev
if true_sev >= _MEDIUM_LOWER:
return 0.80 * true_sev
# Low-medium investigation is barely useful
return 0.40 * true_sev
if action_type == "ESCALATE":
if is_fp or true_sev < _FALSE_POSITIVE_THRESHOLD:
return 0.0
if true_sev >= _HIGH_THRESHOLD:
return 0.85 * true_sev
if true_sev >= _MEDIUM_LOWER:
return 0.60 * true_sev
return 0.30 * true_sev
if action_type == "IGNORE":
if is_fp or true_sev < _FALSE_POSITIVE_THRESHOLD:
# Efficient noise filtering β small bonus
return _FP_BONUS_CAP_PER_ALERT * (1.0 - true_sev)
# Ignoring a non-FP alert gives zero (or negative for criticals,
# tracked separately via critical_missed)
return 0.0
# DELAY β deferred, no score contribution this step
return 0.0
# ---------------------------------------------------------------------------
# Evaluation helper
# ---------------------------------------------------------------------------
def run_episode_evaluation(
agent,
env,
num_episodes: int = 10,
seed_offset: int = 0,
verbose: bool = False,
) -> Dict[str, Any]:
"""
Run multiple episodes and return aggregated grading results.
Args:
agent: Agent with .act(observation) -> Action method.
env: AdaptiveAlertTriageEnv(task_id="medium") instance.
num_episodes: Number of episodes to run.
seed_offset: Added to episode index for the reset seed.
verbose: Print per-episode summary when True.
Returns:
Dict with keys: mean_score, std_score, min_score, max_score,
success_rate, mean_efficiency, episode_scores, episode_metrics.
"""
episode_scores: List[float] = []
episode_metrics: List[Dict[str, Any]] = []
for ep in range(num_episodes):
grader = MediumTaskGrader(max_investigations_per_step=3)
obs = env.reset(seed=seed_offset + ep)
done = False
while not done:
if not obs.alerts:
break
action = agent.act(obs)
obs, _reward, done, info = env.step(action)
for alert_data in info.get("processed_alerts", []):
grader.process_step(alert_data, info)
score = grader.get_episode_score()
metrics = grader.get_metrics()
episode_scores.append(score)
episode_metrics.append(metrics)
if verbose:
print(
f" ep {ep + 1:02d} score={score:.3f} "
f"eff={metrics['resource_efficiency']:.3f} "
f"invest={metrics['total_investigations']} "
f"passed={metrics['passed']}"
)
scores_arr = np.array(episode_scores)
eff_arr = np.array([m["resource_efficiency"] for m in episode_metrics])
return {
"mean_score": float(scores_arr.mean()),
"std_score": float(scores_arr.std()),
"min_score": float(scores_arr.min()),
"max_score": float(scores_arr.max()),
"success_rate": float((scores_arr >= SUCCESS_THRESHOLD).mean()),
"mean_efficiency": float(eff_arr.mean()),
"episode_scores": episode_scores,
"episode_metrics": episode_metrics,
}
# ---------------------------------------------------------------------------
# Self-test
# ---------------------------------------------------------------------------
if __name__ == "__main__":
print("MediumTaskGrader β self-test\n" + "=" * 55)
from adaptive_alert_triage.models import Alert, Action, Reward
def _alert(aid: str, true_sev: float, is_fp: bool = False) -> Alert:
return Alert(
id=aid, visible_severity=0.6, confidence=0.85,
alert_type="CPU", age=1, true_severity=true_sev,
metadata={"false_positive": is_fp},
)
grader = MediumTaskGrader()
cases = [
("Critical + INVESTIGATE (best)", "INVESTIGATE", 0.90, False),
("High + ESCALATE", "ESCALATE", 0.80, False),
("Medium + INVESTIGATE", "INVESTIGATE", 0.60, False),
("FP + IGNORE (efficient)", "IGNORE", 0.15, True),
("FP + INVESTIGATE (wasteful)", "INVESTIGATE", 0.15, True),
("Critical + IGNORE (miss)", "IGNORE", 0.90, False),
("Medium + DELAY", "DELAY", 0.60, False),
]
all_pass = True
for desc, act, sev, is_fp in cases:
alert = _alert("ax", sev, is_fp)
action = Action(alert_id="ax", action_type=act)
contrib = grader.grade_action(action, alert, Reward(value=0.0))
print(f" {desc:45s} contrib={contrib:+.4f}")
score = grader.get_episode_score()
m = grader.get_metrics()
print(f"\nEpisode score : {score:.4f}")
print(f"Passed : {m['passed']}")
print(f"Resource efficiency: {m['resource_efficiency']:.4f}")
print(f"Critical missed : {m['critical_missed']}/{m['critical_total']}")
print(f"Unnecessary invest : {m['unnecessary_invest']}/{m['total_investigations']}")
print(f"Action breakdown : {m['action_breakdown']}") |