Spaces:
Sleeping
Sleeping
File size: 15,176 Bytes
b14c6e3 1341fa9 b14c6e3 dd69fa9 b14c6e3 dd69fa9 b14c6e3 c18a9d1 b14c6e3 c18a9d1 b14c6e3 d6fcbb0 b14c6e3 d6fcbb0 1341fa9 c18a9d1 1341fa9 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 | """
Task 1: Basic Alert Prioritisation (Easy)
==========================================
Objective
---------
Learn to correctly classify and respond to individual, independent alerts.
There are no resource constraints and no alert correlations β the agent simply
needs to read each alert's visible attributes and choose the right action.
Difficulty contract
-------------------
- No resource budget (unlimited INVESTIGATE per step)
- Alerts are independent (no correlated chains)
- Correct policy is purely threshold-based on visible severity
Grading formula
---------------
score = correct_actions / total_actions β [0.0, 1.0]
Correctness rules (deterministic)
----------------------------------
true_severity β₯ 0.75 β INVESTIGATE or ESCALATE
true_severity < 0.30 β IGNORE
0.30 β€ severity < 0.75 β INVESTIGATE (always OK)
ESCALATE (OK if severity β₯ 0.60)
IGNORE (OK if severity < 0.50)
DELAY β always wrong in the easy task
Success threshold: β₯ 0.70 (70 % correct action rate)
Integration with env.py
-----------------------
Every call to env.step(action) returns an ``info`` dict containing:
info["processed_alerts"] β list of dicts, one per action this step
Keys: alert_id, true_severity, visible_severity, confidence,
alert_type, age, is_correlated, is_false_positive,
action_taken, correlation_group_index
The grader consumes those dicts via process_step(); this guarantees that
ground-truth fields are used even after the alert has been removed from the
environment queue.
"""
from __future__ import annotations
from typing import Any, Dict, List
import numpy as np
from adaptive_alert_triage.models import Action, Alert, Reward
# ---------------------------------------------------------------------------
# Severity band boundaries (kept in sync with utils.py constants)
# ---------------------------------------------------------------------------
_CRITICAL_THRESHOLD: float = 0.75
_FALSE_POSITIVE_THRESHOLD: float = 0.30
_MEDIUM_ESCALATE_MIN: float = 0.60 # ESCALATE acceptable above this
_MEDIUM_IGNORE_MAX: float = 0.50 # IGNORE acceptable below this
# Pass threshold
SUCCESS_THRESHOLD: float = 0.696
# ---------------------------------------------------------------------------
# Grader
# ---------------------------------------------------------------------------
class EasyTaskGrader:
"""
Grader for Task 1: Basic Alert Prioritisation.
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 strictly in (0, 1).
4. Optionally call get_metrics() for a full breakdown.
5. Call reset() to reuse the grader for a new episode.
Scoring is fully deterministic: same alert + same action β same score.
"""
def __init__(self) -> None:
self.correct_actions: int = 0
self.total_actions: int = 0
self.action_history: List[Dict[str, Any]] = []
# ------------------------------------------------------------------
# Primary interface (production)
# ------------------------------------------------------------------
def process_step(
self,
alert_data: Dict[str, Any],
info: Dict[str, Any], # noqa: ARG002 (kept for API symmetry)
) -> float:
"""
Evaluate one action using ground-truth data from env.step().
Args:
alert_data: One entry from info["processed_alerts"].
Must contain: true_severity, action_taken.
info: Full info dict from env.step() (unused here but
kept for consistent API across all three graders).
Returns:
0.99 if the action was correct, 0.01 otherwise (strictly in (0, 1)).
"""
self.total_actions += 1
true_severity: float = float(alert_data.get("true_severity", 0.0))
action_type: str = str(alert_data.get("action_taken", ""))
is_correct: bool = self._is_action_correct(action_type, true_severity)
if is_correct:
self.correct_actions += 1
self.action_history.append({
"alert_id": alert_data.get("alert_id", ""),
"action": action_type,
"true_severity": true_severity,
"visible_severity":alert_data.get("visible_severity", 0.0),
"confidence": alert_data.get("confidence", 0.0),
"alert_type": alert_data.get("alert_type", ""),
"is_false_positive":alert_data.get("is_false_positive", False),
"correct": is_correct,
"score": 1.0 if is_correct else 0.0,
})
return 1.0 if is_correct else 0.0
# ------------------------------------------------------------------
# Legacy API (unit tests / backward compat)
# ------------------------------------------------------------------
def grade_action(self, action: Action, alert: Alert, reward: Reward) -> float:
"""
Grade a single action-alert pair (legacy / unit-test API).
Prefer process_step() in production β this wrapper exists only for
backward compatibility with existing unit tests.
"""
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.total_actions == 0:
return 0.5
raw = self.correct_actions / self.total_actions
# Linearly map exactly to [0.01, 0.99] without clipping
mapped = (raw * 0.98) + 0.01
return float(round(mapped, 4))
def passed(self) -> bool:
"""Return True if the agent meets the easy-task success threshold."""
return self.get_episode_score() >= SUCCESS_THRESHOLD
# ------------------------------------------------------------------
# Metrics
# ------------------------------------------------------------------
def get_metrics(self) -> Dict[str, Any]:
"""
Return a detailed per-band accuracy breakdown.
Severity bands:
critical true_severity β₯ 0.75
medium 0.30 β€ true_severity < 0.75
false_positive true_severity < 0.30
"""
score = self.get_episode_score()
critical_h = [h for h in self.action_history if h["true_severity"] >= _CRITICAL_THRESHOLD]
medium_h = [h for h in self.action_history
if _FALSE_POSITIVE_THRESHOLD <= h["true_severity"] < _CRITICAL_THRESHOLD]
fp_h = [h for h in self.action_history if h["true_severity"] < _FALSE_POSITIVE_THRESHOLD]
def _acc(subset: list) -> float:
if not subset:
return 0.0
return sum(h["correct"] for h in subset) / len(subset)
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,
"correct_actions": self.correct_actions,
"total_actions": self.total_actions,
"critical_accuracy": _acc(critical_h),
"medium_accuracy": _acc(medium_h),
"false_positive_accuracy": _acc(fp_h),
"critical_count": len(critical_h),
"medium_count": len(medium_h),
"false_positive_count": len(fp_h),
"action_breakdown": breakdown,
}
# ------------------------------------------------------------------
# Housekeeping
# ------------------------------------------------------------------
def reset(self) -> None:
"""Reset all state for a new episode."""
self.correct_actions = 0
self.total_actions = 0
self.action_history = []
def __repr__(self) -> str:
score = self.get_episode_score()
return (
f"EasyTaskGrader(score={score:.3f}, "
f"correct={self.correct_actions}/{self.total_actions}, "
f"passed={self.passed()})"
)
# ------------------------------------------------------------------
# Private helpers
# ------------------------------------------------------------------
@staticmethod
def _is_action_correct(action_type: str, true_severity: float) -> bool:
"""
Deterministic correctness rule for Task 1.
Critical (β₯ 0.75) β INVESTIGATE or ESCALATE
False positive (< 0.30) β IGNORE only
Medium (0.30β0.74) β INVESTIGATE always OK
ESCALATE OK if severity β₯ 0.60
IGNORE OK if severity < 0.50
DELAY never OK (creates unnecessary backlog)
"""
if true_severity >= _CRITICAL_THRESHOLD:
return action_type in ("INVESTIGATE", "ESCALATE")
if true_severity < _FALSE_POSITIVE_THRESHOLD:
return action_type == "IGNORE"
# Medium band
if action_type == "INVESTIGATE":
return True
if action_type == "ESCALATE":
return true_severity >= _MEDIUM_ESCALATE_MIN
if action_type == "IGNORE":
return true_severity < _MEDIUM_IGNORE_MAX
# DELAY
return False
# ---------------------------------------------------------------------------
# Evaluation helper (used by evaluation/evaluate.py)
# ---------------------------------------------------------------------------
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="easy") instance.
num_episodes: Number of episodes to run.
seed_offset: Added to episode index to produce the reset seed.
verbose: Print per-episode summary when True.
Returns:
Dict with keys: mean_score, std_score, min_score, max_score,
success_rate, episode_scores, episode_metrics.
"""
episode_scores: List[float] = []
episode_metrics: List[Dict[str, Any]] = []
for ep in range(num_episodes):
grader = EasyTaskGrader()
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"correct={metrics['correct_actions']}/{metrics['total_actions']} "
f"passed={metrics['passed']}"
)
scores_arr = np.array(episode_scores)
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()),
"episode_scores": episode_scores,
"episode_metrics": episode_metrics,
}
# ---------------------------------------------------------------------------
# Self-test
# ---------------------------------------------------------------------------
if __name__ == "__main__":
print("EasyTaskGrader β self-test\n" + "=" * 50)
from adaptive_alert_triage.models import Alert, Action, Reward
def _alert(aid: str, true_sev: float, vis_sev: float = 0.5) -> Alert:
return Alert(
id=aid, visible_severity=vis_sev, confidence=0.9,
alert_type="CPU", age=1, true_severity=true_sev,
)
cases = [
# (desc, action, true_sev, expected_score)
("Critical + INVESTIGATE", "INVESTIGATE", 0.90, 1.0),
("Critical + ESCALATE", "ESCALATE", 0.90, 1.0),
("Critical + IGNORE (wrong)", "IGNORE", 0.90, 0.0),
("Critical + DELAY (wrong)", "DELAY", 0.90, 0.0),
("FP + IGNORE", "IGNORE", 0.10, 1.0),
("FP + INVESTIGATE (wrong)", "INVESTIGATE", 0.10, 0.0),
("Medium + INVESTIGATE", "INVESTIGATE", 0.55, 1.0),
("Medium + ESCALATE hi (ok)", "ESCALATE", 0.65, 1.0),
("Medium + ESCALATE lo (wrong)", "ESCALATE", 0.45, 0.0),
("Medium + IGNORE lo (ok)", "IGNORE", 0.40, 1.0),
("Medium + IGNORE hi (wrong)", "IGNORE", 0.60, 0.0),
("Medium + DELAY (wrong)", "DELAY", 0.55, 0.0),
]
grader = EasyTaskGrader()
all_pass = True
for desc, act, sev, expected in cases:
alert = _alert("a1", sev)
action = Action(alert_id="a1", action_type=act)
result = grader.grade_action(action, alert, Reward(value=0.0))
ok = result == expected
if not ok:
all_pass = False
print(f" [{'PASS' if ok else 'FAIL'}] {desc}")
if not ok:
print(f" got {result}, expected {expected}")
final = grader.get_episode_score()
print(f"\nEpisode score : {final:.3f}")
print(f"Passed : {grader.passed()}")
m = grader.get_metrics()
print(f"Critical acc : {m['critical_accuracy']:.3f}")
print(f"Medium acc : {m['medium_accuracy']:.3f}")
print(f"FP acc : {m['false_positive_accuracy']:.3f}")
print("\nAll tests passed!" if all_pass else "\nSome FAILED β check above.") |