Spaces:
Sleeping
Sleeping
File size: 27,083 Bytes
b14c6e3 8061f1b b14c6e3 8061f1b b14c6e3 8061f1b b14c6e3 9c3fa6e b14c6e3 9c3fa6e b14c6e3 9c3fa6e b14c6e3 eea342f b14c6e3 eea342f b14c6e3 eea342f b14c6e3 c18a9d1 b14c6e3 c18a9d1 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 | """
Adaptive Alert Triage Environment β OpenEnv-compliant OpenEnv Environment
Implements a partially observable RL environment that simulates a real-world
DevOps / SOC alert-triage workflow. An agent must process a continuous stream
of system alerts under time and resource constraints, deciding for each alert:
INVESTIGATE β allocate resources to diagnose (costly)
IGNORE β dismiss as noise (efficient for false positives)
ESCALATE β route to specialist team
DELAY β defer to the next time-step
The environment supports three difficulty tasks:
easy (30 steps, no resource constraint, 10 % correlation probability)
medium (40 steps, K=3 investigations/step, 20 % correlation probability)
hard (50 steps, K=3 investigations/step, 40 % correlation probability,
stricter failure threshold)
OpenEnv interface
-----------------
reset(seed?, options?) -> Observation
step(action) -> (Observation, Reward, done, info)
state() -> EpisodeState
Info dict keys (required by graders)
-------------------------------------
processed_alerts : list[dict] β ground-truth data for every action taken
this step (alert_id, true_severity,
is_false_positive, action_taken, etc.)
correlation_groups: list[list] β current correlated-chain groups (alert IDs)
failures_this_step: int β failures triggered this step
system_failure : bool β True if the episode is in a failure state
step : int β current step index
cumulative_reward : float β total reward so far
failures_count : int β total failures so far
action_correct : bool β whether the most recent action was optimal
"""
from __future__ import annotations
from collections import deque
from typing import Any, Dict, List, Optional, Tuple
import numpy as np
import openenv_shim as gym
from openenv_shim import spaces
from adaptive_alert_triage.models import (
Action,
Alert,
EpisodeState,
Observation,
Reward,
)
from adaptive_alert_triage import utils
# Import reward calculation with graceful fallback for development mode
import os as _os
import sys as _sys
try:
from rewards.reward import calculate_reward
except ImportError:
_project_root = _os.path.dirname(
_os.path.dirname(_os.path.dirname(_os.path.abspath(__file__)))
)
if _project_root not in _sys.path:
_sys.path.insert(0, _project_root)
from rewards.reward import calculate_reward # type: ignore[no-redef]
# ---------------------------------------------------------------------------
# Task configurations
# ---------------------------------------------------------------------------
_TASK_CONFIGS: Dict[str, Dict[str, Any]] = {
"easy": {
"max_steps": 10,
"failure_threshold": 2,
"max_investigations": None, # unconstrained
"correlation_probability": 0.10,
"description": "Basic alert prioritisation β no resource constraint.",
},
"medium": {
"max_steps": 15,
"failure_threshold": 3,
"max_investigations": 3, # K = 3 per step
"correlation_probability": 0.20,
"description": "Resource-constrained triage β K=3 investigations/step.",
},
"hard": {
"max_steps": 20,
"failure_threshold": 2, # stricter
"max_investigations": 3,
"correlation_probability": 0.40,
"description": (
"Cascading-failure prevention β correlated alerts, delayed failures, "
"hidden severity, strict failure threshold."
),
},
}
# ---------------------------------------------------------------------------
# Main environment class
# ---------------------------------------------------------------------------
class AdaptiveAlertTriageEnv(gym.Env):
"""
OpenEnv environment for adaptive alert triage.
Parameters
----------
task_id : str
Difficulty level: ``"easy"``, ``"medium"``, or ``"hard"``.
max_steps : int, optional
Override the task-default episode length.
seed : int, optional
Fixed random seed for full reproducibility.
"""
metadata = {"render_modes": ["human", "ansi"]}
# ------------------------------------------------------------------
# Construction
# ------------------------------------------------------------------
def __init__(
self,
task_id: str = "easy",
max_steps: Optional[int] = None,
seed: Optional[int] = None,
) -> None:
super().__init__()
if task_id not in _TASK_CONFIGS:
raise ValueError(
f"Unknown task_id '{task_id}'. "
f"Valid options: {sorted(_TASK_CONFIGS.keys())}"
)
self.task_id: str = task_id
self.config: Dict[str, Any] = dict(_TASK_CONFIGS[task_id])
self.max_steps: int = max_steps or self.config["max_steps"]
self.failure_threshold: int = self.config["failure_threshold"]
self.max_investigations_per_step: Optional[int] = self.config["max_investigations"]
# Episode state β initialised properly in reset()
self.current_step: int = 0
self.alerts: List[Alert] = []
self.failures_count: int = 0
self.cumulative_reward: float = 0.0
self.investigations_used: int = 0
# Hidden state
self.correlation_groups: List[List[str]] = []
# Action history (for state() and checkpointing)
self._action_history: List[Action] = []
# Real-alert ingestion queue (Datadog / Kafka webhook mode)
self.real_alerts_queue: deque = deque(maxlen=50)
# Per-step grading data β populated in step(), consumed by graders
self._processed_alerts_this_step: List[Dict[str, Any]] = []
self._failures_this_step: int = 0
# Seed
self._seed: Optional[int] = seed
if seed is not None:
utils.set_seed(seed)
# OpenEnv spaces (abstract; real actions are Action Pydantic objects)
self.action_space = spaces.Discrete(4) # 4 ActionType values
self.observation_space = spaces.Dict(
{
"system_load": spaces.Box(0.0, 1.0, shape=(1,), dtype=np.float32),
"queue_length": spaces.Box(0, 100, shape=(1,), dtype=np.int32),
"time_remaining": spaces.Box(
0, self.max_steps, shape=(1,), dtype=np.int32
),
}
)
# ------------------------------------------------------------------
# OpenEnv interface β reset
# ------------------------------------------------------------------
def reset(
self,
seed: Optional[int] = None,
options: Optional[Dict[str, Any]] = None,
) -> Observation:
"""
Reset the environment to a clean initial state.
Args:
seed: Override seed for this episode.
options: Reserved for future use (ignored).
Returns:
Initial Observation with no agent-visible hidden fields.
"""
if seed is not None:
self._seed = seed
if self._seed is not None:
utils.set_seed(self._seed)
# Reset all episode counters
self.current_step = 0
self.failures_count = 0
self.cumulative_reward = 0.0
self.investigations_used = 0
self.correlation_groups = []
self._action_history = []
self._processed_alerts_this_step = []
self._failures_this_step = 0
# Generate the initial alert batch
self.alerts = self._generate_initial_alerts()
return self._create_observation()
# ------------------------------------------------------------------
# OpenEnv interface β step
# ------------------------------------------------------------------
def step(
self, action: Action
) -> Tuple[Observation, Reward, bool, Dict[str, Any]]:
"""
Execute one environment step.
The agent submits one Action per call; the environment:
1. Validates the alert ID and resource budget.
2. Records ground-truth data for the graders.
3. Calculates the dense reward.
4. Applies the action (removes / keeps alert).
5. Ages remaining alerts.
6. Checks for delayed failures.
7. Generates new alerts (Poisson arrivals + possible correlation chain).
8. Increments step counter and resets per-step budget.
9. Returns (Observation, Reward, done, info).
The ``info`` dict always contains:
processed_alerts β list of ground-truth dicts, one per action
correlation_groups β current correlation chains
failures_this_step β failures triggered this step
system_failure β whether the system is in a failure state
step β current step index
cumulative_reward β total reward this episode
failures_count β total failures this episode
action_correct β whether the action matched the optimal policy
Args:
action: Agent's Action targeting one alert by ID.
Returns:
(next_observation, reward, done, info)
"""
# --- Reset per-step tracking ---
self._processed_alerts_this_step = []
self._failures_this_step = 0
# --- Validate alert ID ---
alert = self._get_alert_by_id(action.alert_id)
if alert is None:
reward = Reward(
value=-5.0,
components={"invalid_action": -5.0},
info={"error": f"Alert ID '{action.alert_id}' not found in queue"},
)
obs = self._create_observation()
return obs, reward, True, self._build_info(
action_correct=False,
extra={"error": "Invalid alert ID"},
)
# --- Resource-budget enforcement ---
if (
self.max_investigations_per_step is not None
and action.action_type == "INVESTIGATE"
):
if self.investigations_used >= self.max_investigations_per_step:
reward = Reward(
value=-3.0,
components={"resource_budget_exceeded": -3.0},
info={
"error": "Investigation budget exhausted for this step",
"budget": self.max_investigations_per_step,
"used": self.investigations_used,
},
)
obs = self._create_observation()
return obs, reward, False, self._build_info(
action_correct=False,
extra={"resource_constraint_violated": True},
)
self.investigations_used += 1
# --- Record ground-truth for graders BEFORE removing the alert ---
processed: Dict[str, Any] = {
"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,
"is_correlated": alert.is_correlated,
"is_false_positive": bool(alert.metadata.get("false_positive", False)),
"action_taken": action.action_type,
"correlation_group_index": self._find_correlation_group(alert.id),
}
self._processed_alerts_this_step.append(processed)
# --- Track action history ---
self._action_history.append(action)
# --- Calculate dense reward ---
reward = calculate_reward(action, alert, self.config)
self.cumulative_reward += reward.value
# --- Apply action to alert queue ---
self._process_action(action, alert)
# --- Age all remaining unresolved alerts ---
self._age_alerts()
# --- Check for failures triggered by aged critical alerts ---
self._failures_this_step = self._check_for_failures()
self.failures_count += self._failures_this_step
# --- Generate new alerts (Poisson arrivals + possible chain) ---
if utils.should_generate_new_alerts(self.current_step, len(self.alerts)):
new_alerts = self._generate_new_alerts()
self.alerts.extend(new_alerts)
# --- Advance step and reset per-step investigation budget ---
self.current_step += 1
self.investigations_used = 0
# --- Termination check ---
done: bool = self._is_terminal()
# --- Build next observation (hidden fields masked) ---
obs = self._create_observation()
# --- Determine overall failure state ---
system_in_failure: bool = (
self.failures_count >= self.failure_threshold
or self._failures_this_step > 0
)
info = self._build_info(
action_correct=bool(reward.info.get("action_correct", False)),
extra={
"system_failure": system_in_failure,
"alert_handled": alert.id,
},
)
return obs, reward, done, info
# ------------------------------------------------------------------
# OpenEnv interface β state
# ------------------------------------------------------------------
def state(self) -> EpisodeState:
"""
Return the complete internal episode state (visible + hidden).
Used by evaluation scripts, replay tools, and the hard-task grader
for root-cause analysis. NOT intended to be passed to the agent.
Returns:
EpisodeState with full ground-truth information.
"""
hidden: Dict[str, Any] = {
"true_severities": {a.id: a.true_severity for a in self.alerts},
"correlation_groups": [list(g) for g in self.correlation_groups],
"false_positives": [
a.id for a in self.alerts
if a.metadata.get("false_positive", False)
],
# Pending failures: alerts that are critical AND close to the age threshold
"pending_failures": {
a.id: utils.CRITICAL_AGE_THRESHOLD - a.age
for a in self.alerts
if utils.is_critical_alert(a)
and a.age < utils.CRITICAL_AGE_THRESHOLD
},
}
return EpisodeState(
observation=self._create_observation(),
hidden_state=hidden,
cumulative_reward=self.cumulative_reward,
failures_count=self.failures_count,
actions_taken=[a.model_dump() for a in self._action_history],
seed=self._seed,
)
# ------------------------------------------------------------------
# Internal helpers β observation construction
# ------------------------------------------------------------------
def _create_observation(self) -> Observation:
"""
Build the agent-facing Observation by masking all hidden fields.
true_severity and is_correlated are zeroed-out; metadata is stripped.
The agent must infer hidden information from visible_severity,
confidence, alert_type, and age alone.
"""
system_load: float = utils.calculate_system_load(len(self.alerts))
visible_alerts: List[Alert] = []
for a in self.alerts:
visible_alerts.append(
Alert(
id=a.id,
visible_severity=a.visible_severity,
confidence=a.confidence,
alert_type=a.alert_type,
age=a.age,
# Hidden fields zeroed out
true_severity=0.0,
is_correlated=False,
metadata={},
)
)
resource_budget: Optional[int] = None
if self.max_investigations_per_step is not None:
resource_budget = self.max_investigations_per_step - self.investigations_used
return Observation(
alerts=visible_alerts,
system_load=system_load,
queue_length=len(self.alerts),
time_remaining=max(0, self.max_steps - self.current_step),
episode_step=self.current_step,
resource_budget=resource_budget,
)
# ------------------------------------------------------------------
# Internal helpers β alert generation
# ------------------------------------------------------------------
def _generate_initial_alerts(self) -> List[Alert]:
"""
Generate the starting alert batch for a fresh episode.
Real alerts from the ingestion queue are prioritised; any remaining
slots are filled with synthetic alerts.
"""
num_initial: int = int(np.random.randint(3, 7))
alerts: List[Alert] = []
# Drain real alerts first
while self.real_alerts_queue and len(alerts) < num_initial:
raw = self.real_alerts_queue.popleft()
alerts.append(self._ingest_real_alert(raw))
# Fill with synthetic
for i in range(len(alerts), num_initial):
alerts.append(
utils.generate_alert(step=0, alert_index=i)
)
return alerts
def _generate_new_alerts(self) -> List[Alert]:
"""
Generate alerts to append to the queue this step.
If real alerts are queued they are processed first (no synthetic
alerts generated that step). Otherwise, a Poisson-sampled batch of
independent alerts is generated, with a task-configured probability
that a correlated chain replaces the batch entirely.
"""
# Priority: real ingest queue
if self.real_alerts_queue:
raw = self.real_alerts_queue.popleft()
return [self._ingest_real_alert(raw)]
# Correlated chain vs independent batch
if np.random.random() < self.config["correlation_probability"]:
chain_alerts = utils.generate_correlated_alerts(
self.current_step, num_alerts=3
)
self.correlation_groups.append([a.id for a in chain_alerts])
return chain_alerts
num_new: int = utils.sample_num_new_alerts()
return [
utils.generate_alert(
step=self.current_step,
alert_index=i,
)
for i in range(num_new)
]
@staticmethod
def _ingest_real_alert(raw: Dict[str, Any]) -> Alert:
"""
Convert a raw real-alert dict into an Alert with synthesised ground truth.
Ground truth is estimated by adding Gaussian noise to visible_severity,
reflecting that real monitoring tools provide imperfect severity scores.
"""
true_severity: float = float(
np.clip(
float(raw["visible_severity"]) + np.random.normal(0.0, 0.10),
0.0,
1.0,
)
)
return Alert(
id=raw["id"],
visible_severity=float(raw["visible_severity"]),
confidence=float(raw["confidence"]),
alert_type=raw["type"],
age=0,
true_severity=true_severity,
is_correlated=False,
metadata={"source": "real_ingest", "raw": raw},
)
# ------------------------------------------------------------------
# Internal helpers β action processing
# ------------------------------------------------------------------
def _process_action(self, action: Action, alert: Alert) -> None:
"""
Apply the agent's action to the alert queue.
INVESTIGATE, ESCALATE, and IGNORE all resolve the alert (remove it
from the queue). DELAY keeps the alert in the queue; its age will
be incremented by _age_alerts().
"""
if action.action_type in ("INVESTIGATE", "ESCALATE", "IGNORE"):
self.alerts = [a for a in self.alerts if a.id != alert.id]
# DELAY: no-op β alert remains; age increment handled in _age_alerts()
def _age_alerts(self) -> None:
"""Increment the age of every unresolved alert by one step."""
for alert in self.alerts:
alert.age += 1
def _check_for_failures(self) -> int:
"""
Detect and remove alerts that have caused system failures.
A failure occurs when a critical alert (true_severity β₯ 0.75) has
been in the queue for CRITICAL_AGE_THRESHOLD or more steps without
being resolved. Each such alert contributes one failure event.
Returns:
Number of new failure events detected this step.
"""
failures: int = 0
failed_ids: List[str] = []
for alert in self.alerts:
if (
utils.is_critical_alert(alert)
and alert.age >= utils.CRITICAL_AGE_THRESHOLD
):
failures += 1
failed_ids.append(alert.id)
# Remove failed alerts (they've escalated out of the triage queue)
if failed_ids:
self.alerts = [a for a in self.alerts if a.id not in failed_ids]
return failures
# ------------------------------------------------------------------
# Internal helpers β utilities
# ------------------------------------------------------------------
def _get_alert_by_id(self, alert_id: str) -> Optional[Alert]:
"""Return the Alert with the given ID, or None if not found."""
for alert in self.alerts:
if alert.id == alert_id:
return alert
return None
def _find_correlation_group(self, alert_id: str) -> Optional[int]:
"""
Return the index of the correlation group that contains alert_id, or None.
Used to populate the ``correlation_group_index`` field in processed_alerts
so the hard-task grader can score root-cause identification.
"""
for idx, group in enumerate(self.correlation_groups):
if alert_id in group:
return idx
return None
def _is_terminal(self) -> bool:
"""Return True if the episode should end."""
return (
self.current_step >= self.max_steps
or self.failures_count >= self.failure_threshold
)
def _build_info(
self,
action_correct: bool,
extra: Optional[Dict[str, Any]] = None,
) -> Dict[str, Any]:
"""
Assemble the standard info dict returned from step().
Always includes the keys required by all three task graders.
Additional keys can be merged in via ``extra``.
"""
info: Dict[str, Any] = {
# Core grading keys (required)
"processed_alerts": list(self._processed_alerts_this_step),
"correlation_groups": [list(g) for g in self.correlation_groups],
"failures_this_step": self._failures_this_step,
"system_failure": self.failures_count >= self.failure_threshold
or self._failures_this_step > 0,
# Convenience telemetry
"step": self.current_step,
"cumulative_reward": self.cumulative_reward,
"failures_count": self.failures_count,
"action_correct": action_correct,
}
if extra:
info.update(extra)
return info
# ------------------------------------------------------------------
# OpenEnv render
# ------------------------------------------------------------------
def render(self, mode: str = "human") -> Optional[str]:
"""
Render a text summary of the current environment state.
Args:
mode: ``"human"`` (prints to stdout) or ``"ansi"`` (returns string).
Returns:
String if mode is ``"ansi"``, otherwise None.
"""
lines = [
f"\n=== Step {self.current_step}/{self.max_steps}"
f" [{self.task_id}] ===",
f" Failures : {self.failures_count}/{self.failure_threshold}",
f" Cum. reward : {self.cumulative_reward:+.1f}",
f" Active alerts: {len(self.alerts)}",
]
if self.max_investigations_per_step is not None:
lines.append(
f" Inv. budget : "
f"{self.max_investigations_per_step - self.investigations_used}"
f"/{self.max_investigations_per_step} remaining"
)
if self.alerts:
lines.append("\n Alerts (first 5):")
for a in self.alerts[:5]:
lines.append(
f" {a.id} sev={a.visible_severity:.2f}"
f" conf={a.confidence:.2f}"
f" type={a.alert_type:<12}"
f" age={a.age}"
)
if len(self.alerts) > 5:
lines.append(f" β¦ and {len(self.alerts) - 5} more")
output = "\n".join(lines) + "\n"
if mode == "human":
print(output)
return None
return output
# ---------------------------------------------------------------------------
# Quick demo
# ---------------------------------------------------------------------------
def main() -> None:
"""Run a short demo episode with a simple heuristic policy."""
print("Adaptive Alert Triage Environment β Demo\n")
env = AdaptiveAlertTriageEnv(task_id="easy", seed=42)
obs: Observation = env.reset()
print(f"Initial observation: {len(obs.alerts)} alerts "
f"(system_load={obs.system_load:.2f})\n")
done = False
step_count = 0
while not done and step_count < 5:
env.render()
if not obs.alerts:
print("No alerts in queue β nothing to handle.")
break
# Heuristic: pick the alert with the highest visible_severity
best_alert = max(obs.alerts, key=lambda a: a.visible_severity)
action = Action(
alert_id=best_alert.id,
action_type=(
"INVESTIGATE" if best_alert.visible_severity >= 0.7 else "IGNORE"
),
)
obs, reward, done, info = env.step(action)
print(
f" Action: {action.action_type} β {best_alert.id}"
f" Reward: {reward.value:+.1f}"
f" Correct: {info.get('action_correct', '?')}"
)
step_count += 1
print(f"\nDemo finished after {step_count} steps.")
print(f"Final cumulative reward : {env.cumulative_reward:+.1f}")
print(f"Total system failures : {env.failures_count}")
if __name__ == "__main__":
main() |