Spaces:
Sleeping
Sleeping
File size: 11,776 Bytes
b14c6e3 c18a9d1 eea342f b14c6e3 9c3fa6e 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 | """
Pydantic Models for OpenEnv Compliance
Defines the Observation, Action, and Reward models using Pydantic for
validation and serialization. These models form the core API contract
for the Adaptive Alert Triage environment.
OpenEnv compliance:
- Observation, Action, Reward are all Pydantic BaseModel subclasses
- step(action) -> (Observation, Reward, bool, dict)
- reset() -> Observation
- state() -> EpisodeState
"""
from typing import List, Literal, Optional, Dict, Any
from pydantic import BaseModel, Field, field_validator, model_validator
# ---------------------------------------------------------------------------
# Type aliases
# ---------------------------------------------------------------------------
AlertType = Literal["CPU", "MEMORY", "DISK", "NETWORK", "APPLICATION", "SECURITY"]
ActionType = Literal["INVESTIGATE", "IGNORE", "ESCALATE", "DELAY"]
# ---------------------------------------------------------------------------
# Alert
# ---------------------------------------------------------------------------
class Alert(BaseModel):
"""
Represents a single alert in the system.
Visible fields are exposed to the agent via Observation.
Hidden fields (true_severity, is_correlated) are only used internally
by the environment for reward calculation and failure checking — they
are zeroed-out / masked before being returned to the agent.
Attributes:
id: Unique alert identifier.
visible_severity: Noisy, observable severity score in [0.0, 1.0].
confidence: Detection confidence level in [0.0, 1.0].
alert_type: Category of the alert (CPU, MEMORY, etc.).
age: Number of time-steps since the alert was generated.
true_severity: Ground-truth severity (hidden from agent).
is_correlated: Whether the alert is part of a correlated failure
chain (hidden from agent).
metadata: Optional key/value context bag.
"""
id: str = Field(..., description="Unique alert identifier")
visible_severity: float = Field(
..., ge=0.0, le=1.0, description="Observable severity score (noisy)"
)
confidence: float = Field(
..., ge=0.0, le=1.0, description="Detection confidence level"
)
alert_type: AlertType = Field(..., description="Alert category")
age: int = Field(..., ge=0, description="Time-steps since generation")
# Hidden attributes — never returned to the agent in Observation
true_severity: float = Field(
default=0.0,
ge=0.0,
le=1.0,
description="Ground-truth severity (hidden from agent)",
)
is_correlated: bool = Field(
default=False,
description="Part of a correlated failure chain (hidden from agent)",
)
metadata: Dict[str, Any] = Field(
default_factory=dict, description="Optional context bag"
)
@field_validator("visible_severity", "confidence", "true_severity", mode="before")
@classmethod
def clamp_to_unit_interval(cls, v: float) -> float:
"""Silently clamp float fields to [0.0, 1.0] to absorb small fp drift."""
return float(max(0.0, min(1.0, float(v))))
model_config = {
"json_schema_extra": {
"example": {
"id": "alert_0001_00",
"visible_severity": 0.82,
"confidence": 0.91,
"alert_type": "CPU",
"age": 1,
}
}
}
# ---------------------------------------------------------------------------
# Observation (what the agent sees each step)
# ---------------------------------------------------------------------------
class Observation(BaseModel):
"""
Environment observation returned by reset() and step().
Contains only the information that is visible to the agent.
Hidden fields (true_severity, is_correlated) are stripped before
this object is constructed — see AdaptiveAlertTriageEnv._create_observation().
Attributes:
alerts: Active alerts requiring triage decisions.
system_load: Current infrastructure utilisation in [0.0, 1.0].
queue_length: Total alerts currently in the processing queue.
time_remaining: Steps left before episode ends.
episode_step: Current step index (0-based).
resource_budget: Remaining investigation actions this step (None if
task has no resource constraint).
"""
alerts: List[Alert] = Field(default_factory=list, description="Active alerts")
system_load: float = Field(
..., ge=0.0, le=1.0, description="System resource utilisation"
)
queue_length: int = Field(..., ge=0, description="Alerts currently in queue")
time_remaining: int = Field(..., ge=0, description="Steps left in episode")
episode_step: int = Field(..., ge=0, description="Current step (0-based)")
resource_budget: Optional[int] = Field(
None,
description="Remaining INVESTIGATE actions this step (None = unconstrained)",
)
@model_validator(mode="after")
def queue_length_matches_alerts(self) -> "Observation":
"""queue_length must equal len(alerts) for consistency."""
if self.queue_length != len(self.alerts):
# Auto-correct rather than raise — keeps the server robust
object.__setattr__(self, "queue_length", len(self.alerts))
return self
model_config = {
"json_schema_extra": {
"example": {
"alerts": [
{
"id": "alert_0003_01",
"visible_severity": 0.85,
"confidence": 0.92,
"alert_type": "CPU",
"age": 3,
}
],
"system_load": 0.72,
"queue_length": 1,
"time_remaining": 47,
"episode_step": 3,
"resource_budget": 2,
}
}
}
# ---------------------------------------------------------------------------
# Action (what the agent sends each step)
# ---------------------------------------------------------------------------
class Action(BaseModel):
"""
Agent action targeting a single alert.
The environment processes one Action per call to step(). Resource-
constrained tasks (medium / hard) count INVESTIGATE actions against a
per-step budget tracked in Observation.resource_budget.
Attributes:
alert_id: ID of the alert to act upon. Must match an id in the
current Observation.alerts list.
action_type: Decision to apply:
INVESTIGATE — deep-dive, costs resource budget.
IGNORE — dismiss as noise (zero cost).
ESCALATE — route to specialist (medium cost).
DELAY — keep in queue, re-evaluate next step.
metadata: Optional free-form context (e.g. escalation reason).
"""
alert_id: str = Field(..., description="Target alert identifier")
action_type: ActionType = Field(..., description="Action to perform")
metadata: Dict[str, Any] = Field(
default_factory=dict, description="Optional action context"
)
model_config = {
"json_schema_extra": {
"example": {
"alert_id": "alert_0003_01",
"action_type": "INVESTIGATE",
"metadata": {"reason": "high visible severity"},
}
}
}
# ---------------------------------------------------------------------------
# Reward (what step() returns alongside the next Observation)
# ---------------------------------------------------------------------------
class Reward(BaseModel):
"""
Scalar reward signal returned after each action.
The dense reward is decomposed into named components so that graders,
evaluation scripts, and debugging tools can inspect each contribution.
Reward schedule:
+10 : Critical alert correctly handled (INVESTIGATE or ESCALATE).
+5 : Action prevented a future cascading failure.
+3 : False positive correctly ignored.
-2 : Unnecessary investigation (benign alert investigated).
-8 : Critical alert missed (IGNORE or excessive DELAY).
-10 : System failure triggered by accumulated unhandled alerts.
Attributes:
value: Total scalar reward for this step.
components: Breakdown mapping component name -> float contribution.
info: Debugging / logging extras (ground-truth reveal, etc.).
"""
value: float = Field(..., description="Total scalar reward")
components: Dict[str, float] = Field(
default_factory=dict, description="Per-component reward breakdown"
)
info: Dict[str, Any] = Field(
default_factory=dict, description="Debugging / ground-truth context"
)
model_config = {
"json_schema_extra": {
"example": {
"value": 10.0,
"components": {
"critical_handled": 10.0,
"false_positive_penalty": 0.0,
"resource_penalty": 0.0,
"failure_penalty": 0.0,
},
"info": {
"alert_id": "alert_0003_01",
"true_severity": 0.90,
"action_correct": True,
"was_false_positive": False,
"was_critical": True,
},
}
}
}
# ---------------------------------------------------------------------------
# EpisodeState (full internal state — used by state() and checkpointing)
# ---------------------------------------------------------------------------
class EpisodeState(BaseModel):
"""
Complete internal snapshot of a running episode.
Returned by AdaptiveAlertTriageEnv.state(). Contains both the visible
observation AND the hidden ground-truth information — intended for
evaluation scripts, replay, and debugging only (never exposed to the
agent during training).
Attributes:
observation: Current agent-visible observation.
hidden_state: Ground-truth data not exposed to the agent:
true_severities — {alert_id: float}
correlation_groups — [[alert_id, ...], ...]
false_positives — [alert_id, ...]
pending_failures — {alert_id: steps_until_failure}
cumulative_reward: Total reward accumulated so far this episode.
failures_count: Number of system failures that have occurred.
actions_taken: Ordered history of every Action in this episode.
seed: Random seed used to initialise this episode.
"""
observation: Observation = Field(..., description="Current agent-visible state")
hidden_state: Dict[str, Any] = Field(
default_factory=dict,
description="Ground-truth data (not exposed to agent)",
)
cumulative_reward: float = Field(
default=0.0, description="Total reward accumulated this episode"
)
failures_count: int = Field(
default=0, ge=0, description="System failures so far"
)
actions_taken: List[Dict[str, Any]] = Field(
default_factory=list, description="Full action history for this episode"
)
seed: Optional[int] = Field(
None, description="Random seed for episode reproducibility"
)
model_config = {"arbitrary_types_allowed": True} |