Spaces:
Sleeping
Sleeping
File size: 25,745 Bytes
a74cbe6 | 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 | from __future__ import annotations
import random
import re
import sys
import uuid
from pathlib import Path
from typing import Any, Dict, List, Optional
# Allow running from repo root or server/
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from openenv.core import Environment
from models import (
AetherTaskFlowAction,
AetherTaskFlowObservation,
AetherTaskFlowState,
ActionType,
)
from env.tasks import generate_tasks, get_profile, apply_dynamic_updates
from env.grader import grade
class AetherTaskFlowEnvironment(Environment):
"""
AETHER-TaskFlow: Adaptive Workflow Management RL Environment.
The agent manages a dynamic task queue under resource constraints
and system uncertainty. Three task scenarios of increasing difficulty
test baseline reasoning, adaptation, and robustness.
"""
SUPPORTS_CONCURRENT_SESSIONS = True
DEFAULT_SEED = 42
def __init__(self, difficulty: str = "easy", default_seed: int = DEFAULT_SEED) -> None:
super().__init__()
if difficulty not in ("easy", "medium", "hard"):
raise ValueError(f"difficulty must be easy/medium/hard, got '{difficulty}'")
self._difficulty = difficulty
self._profile = get_profile(difficulty)
self._default_seed = int(default_seed)
self._state: AetherTaskFlowState = AetherTaskFlowState()
self._rng = random.Random()
self._deferred_tasks: List[Dict[str, Any]] = []
def reset(
self,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
**kwargs: Any,
) -> AetherTaskFlowObservation:
self._reset_rubric()
seed = self._default_seed if seed is None else int(seed)
# === CRITICAL: Ensure determinism ===
random.seed(seed)
self._rng = random.Random(seed)
ep_id = episode_id or str(uuid.uuid4())
profile = self._profile
tasks = generate_tasks(self._difficulty, seed=seed)
task_dicts = [t.to_dict() for t in tasks]
resources = {
"time": float(profile["initial_time"]),
"energy": float(profile["initial_energy"]),
"budget": float(profile["initial_budget"]),
}
self._state = AetherTaskFlowState(
episode_id=ep_id,
step_count=0,
difficulty=self._difficulty,
tasks=task_dicts,
completed_tasks=[],
failed_tasks=[],
deferred_tasks=[],
resources=dict(resources),
initial_resources=dict(resources),
system_health=1.0,
cumulative_value=0.0,
cumulative_reward=0.0,
tasks_completed=0,
tasks_failed=0,
episode_done=False,
seed=seed,
)
self._deferred_tasks = []
self._sync_state_queues()
return self._build_obs(
last_action_type=None,
last_action_task_id=None,
last_action_outcome="Episode started. Select a task to act on.",
reward=0.0,
done=False,
)
def step(
self,
action: AetherTaskFlowAction | Dict[str, Any] | str,
timeout_s: Optional[float] = None,
**kwargs: Any,
) -> AetherTaskFlowObservation:
if self._state.episode_id is None:
self.reset(seed=self._default_seed)
try:
parsed_action = self._coerce_action(action)
return self._step_impl(parsed_action)
except Exception as exc:
return self._safe_step_failure(action, exc)
def _step_impl(self, action: AetherTaskFlowAction) -> AetherTaskFlowObservation:
s = self._state
if s.episode_done:
return self._build_obs(
last_action_type=None,
last_action_task_id=None,
last_action_outcome="Episode already finished.",
reward=0.0,
done=True,
)
s.step_count += 1
profile = self._profile
max_steps: int = profile["max_steps"]
# ---- Apply dynamic task updates (medium/hard) ----
if self._difficulty in ("medium", "hard"):
from env.tasks import apply_dynamic_updates as _upd
task_objs_updated = _upd(
[self._make_task_info(t) for t in s.tasks],
s.step_count,
self._difficulty,
self._rng,
)
s.tasks = [t.to_dict() for t in task_objs_updated]
# ---- Deadline expiry check (before acting) ----
still_alive, newly_failed = [], []
for t in s.tasks:
if t.get("deadline", 1) <= 0 and t["status"] == "pending":
t["status"] = "failed"
newly_failed.append(t)
s.system_health = max(0.0, s.system_health - 0.05)
else:
still_alive.append(t)
s.tasks = still_alive
s.failed_tasks.extend(newly_failed)
s.tasks_failed += len(newly_failed)
# ---- Find the target task ----
task = self._find_task(action.task_id, s.tasks)
if task is None:
# Try deferred list
task = self._find_task(action.task_id, self._deferred_tasks)
raw_reward = 0.0
outcome = ""
if task is None:
raw_reward = -0.5
outcome = (
f"Task {action.task_id} not found in active queue. "
"Choose a valid task_id from the observation."
)
s.system_health = max(0.0, s.system_health - 0.02)
else:
raw_reward, outcome = self._execute_action(action.action_type, task, s, max_steps)
s.cumulative_reward += raw_reward
# ---- Decrement deadlines each step ----
for t in s.tasks:
if t["status"] == "pending":
t["deadline"] = max(0, t["deadline"] - 1)
# ---- Recycle deferred tasks if resources improve ----
from env.algorithms import AWFROX
recycler = AWFROX()
resources_dict = {
"energy": s.resources["energy"],
"budget": s.resources["budget"],
}
active_updated, still_deferred = recycler.recycle_deferred(
s.tasks, self._deferred_tasks, resources_dict, s.step_count
)
s.tasks = active_updated
self._deferred_tasks = still_deferred
self._sync_state_queues()
# ---- Done condition ----
no_more_tasks = len(s.tasks) == 0 and len(self._deferred_tasks) == 0
out_of_time = s.step_count >= max_steps
out_of_resources = (
s.resources["energy"] <= 0 or s.resources["time"] <= 0
)
system_collapse = s.system_health <= 0.0
done = no_more_tasks or out_of_time or out_of_resources or system_collapse
s.episode_done = done
self._sync_state_queues()
return self._build_obs(
last_action_type=action.action_type.value,
last_action_task_id=action.task_id,
last_action_outcome=outcome,
reward=self._normalize_step_reward(raw_reward),
done=done,
)
def message_to_action(self, message: str) -> AetherTaskFlowAction:
"""Convert free-form UI text into a valid environment action."""
return self._coerce_action(message)
def _coerce_action(
self,
action: AetherTaskFlowAction | Dict[str, Any] | str | None,
) -> AetherTaskFlowAction:
if isinstance(action, AetherTaskFlowAction):
return action
if action is None:
return self._recommended_action("No action provided; selected a safe default.")
if isinstance(action, str):
return self._parse_action_message(action)
if isinstance(action, dict):
if "message" in action and isinstance(action["message"], str):
return self._parse_action_message(action["message"])
if "input" in action and isinstance(action["input"], str):
return self._parse_action_message(action["input"])
if "action" in action:
nested_action = action["action"]
if isinstance(nested_action, (dict, str)) or nested_action is None:
return self._coerce_action(nested_action)
recommended = self._recommended_action("Filled missing action fields from the current state.")
normalized_payload = {
"action_type": action.get("action_type", recommended.action_type.value),
"task_id": action.get("task_id", recommended.task_id),
"reasoning": action.get("reasoning", recommended.reasoning),
}
return AetherTaskFlowAction.model_validate(normalized_payload)
raise TypeError(f"Unsupported action input: {type(action)!r}")
def _parse_action_message(self, message: str) -> AetherTaskFlowAction:
normalized = (message or "").strip().lower()
recommended = self._recommended_action(
"Selected the top-ranked task from the current observation."
)
if not normalized:
return recommended
keyword_map = (
(ActionType.OPTIMIZE, ("optimize", "optimise", "tune", "analyze", "analyse")),
(ActionType.DELEGATE, ("delegate", "assign", "handoff", "hand off", "offload")),
(ActionType.DEFER, ("defer", "later", "wait", "skip", "postpone")),
(ActionType.EXECUTE, ("execute", "run", "do", "complete", "process", "start")),
)
chosen_action = recommended.action_type
for action_type, keywords in keyword_map:
if any(keyword in normalized for keyword in keywords):
chosen_action = action_type
break
requested_task_id = self._extract_task_id(normalized)
if requested_task_id is not None and self._task_exists(requested_task_id):
task_id = requested_task_id
else:
task_id = recommended.task_id
return AetherTaskFlowAction(
action_type=chosen_action,
task_id=task_id,
reasoning=f"parsed from '{message.strip()[:80]}'",
)
def _recommended_action(self, reasoning: str) -> AetherTaskFlowAction:
candidates = self._iter_candidate_tasks()
if not candidates:
return AetherTaskFlowAction(
action_type=ActionType.DEFER,
task_id=0,
reasoning=reasoning,
)
from env.algorithms import AETHER, RAPTOR
resources = {
"energy": self._state.resources.get("energy", 0.0),
"budget": self._state.resources.get("budget", 0.0),
"time": self._state.resources.get("time", 0.0),
}
max_steps = self._profile["max_steps"]
ranked = AETHER().rank_tasks(candidates, resources, self._state.step_count, max_steps)
best_task_id, _ = ranked[0]
best_task = next(task for task in candidates if task["task_id"] == best_task_id)
action_type = ActionType(
RAPTOR().decide(best_task, resources, self._state.step_count, max_steps)
)
return AetherTaskFlowAction(
action_type=action_type,
task_id=best_task_id,
reasoning=reasoning,
)
def _iter_candidate_tasks(self) -> List[Dict[str, Any]]:
active_tasks = [task for task in self._state.tasks if task.get("status") == "pending"]
if active_tasks:
return active_tasks
deferred_tasks = [
task for task in self._deferred_tasks if task.get("status") in ("pending", "deferred")
]
return deferred_tasks
def _task_exists(self, task_id: int) -> bool:
return self._find_task(task_id, self._state.tasks) is not None or self._find_task(
task_id, self._deferred_tasks
) is not None
def _extract_task_id(self, text: str) -> Optional[int]:
explicit_match = re.search(r"(?:task|id|#)\s*(\d+)", text)
if explicit_match:
return int(explicit_match.group(1))
loose_match = re.search(r"\b(\d+)\b", text)
if loose_match:
return int(loose_match.group(1))
return None
def _safe_step_failure(
self,
action: AetherTaskFlowAction | Dict[str, Any] | str,
exc: Exception,
) -> AetherTaskFlowObservation:
self._state.episode_done = True
self._state.system_health = max(0.0, self._state.system_health - 0.1)
self._sync_state_queues()
last_action_type = None
last_action_task_id = None
if isinstance(action, AetherTaskFlowAction):
last_action_type = action.action_type.value
last_action_task_id = action.task_id
elif isinstance(action, dict):
raw_action_type = action.get("action_type")
if isinstance(raw_action_type, str):
last_action_type = raw_action_type
raw_task_id = action.get("task_id")
if isinstance(raw_task_id, int):
last_action_task_id = raw_task_id
return self._build_obs(
last_action_type=last_action_type,
last_action_task_id=last_action_task_id,
last_action_outcome=(
f"Step failed safely: {type(exc).__name__}: {str(exc)[:160]}"
),
reward=self._normalize_step_reward(-1.0),
done=True,
)
def _reset_rubric(self) -> None:
"""Called at the start of every reset() — OpenEnv lifecycle hook."""
# No persistent rubric state in this env; this hook satisfies the
# openenv.core.Environment base-class interface.
pass
def get_metadata(self) -> dict:
"""Return environment metadata (used by WebInterfaceManager on startup)."""
return {
"name": "aether_taskflow",
"description": (
"AETHER-TaskFlow: Adaptive Workflow Management RL Environment. "
"Agent manages a dynamic task queue under resource constraints, "
"uncertainty, and time pressure. Real-world enterprise tasks."
),
"difficulty": self._difficulty,
"max_steps": self._profile["max_steps"],
"action_types": ["execute", "defer", "delegate", "optimize"],
"version": "1.0.0",
}
def close(self) -> None:
"""Clean up environment resources (no-op for this in-memory env)."""
pass
# ------------------------------------------------------------------
# state property (OpenEnv required)
# ------------------------------------------------------------------
@property
def state(self) -> AetherTaskFlowState:
self._sync_state_queues()
return self._state
def _execute_action(
self,
action_type: ActionType,
task: Dict[str, Any],
s: AetherTaskFlowState,
max_steps: int,
) -> tuple[float, str]:
"""Execute the chosen action on a task. Returns (reward, outcome_str)."""
energy_cost = task["required_energy"]
budget_cost = task["required_budget"]
uncertainty = task["uncertainty"]
value = task["value"]
priority = task["priority"]
deadline = task["deadline"]
time_left = max_steps - s.step_count
if action_type == ActionType.EXECUTE:
# Check resource sufficiency
if s.resources["energy"] < energy_cost or s.resources["budget"] < budget_cost:
s.system_health = max(0.0, s.system_health - 0.08)
return -1.0, (
f"Cannot execute '{task['name']}': insufficient resources "
f"(need E={energy_cost:.1f}/B={budget_cost:.1f}, "
f"have E={s.resources['energy']:.1f}/B={s.resources['budget']:.1f})."
)
# Uncertainty-based failure chance
success_prob = 1.0 - uncertainty * 0.4
if self._rng.random() > success_prob:
# Partial failure — lose resources, get partial reward
s.resources["energy"] = max(0.0, s.resources["energy"] - energy_cost * 0.5)
s.resources["budget"] = max(0.0, s.resources["budget"] - budget_cost * 0.5)
s.system_health = max(0.0, s.system_health - 0.06)
self._remove_task(task["task_id"], s)
partial_reward = value * priority * 0.25
s.cumulative_value += partial_reward
return partial_reward, (
f"Partial failure on '{task['name']}' (uncertainty={uncertainty:.2f}). "
f"Partial reward: {partial_reward:.2f}."
)
# Success
s.resources["energy"] = max(0.0, s.resources["energy"] - energy_cost)
s.resources["budget"] = max(0.0, s.resources["budget"] - budget_cost)
s.resources["time"] = max(0.0, s.resources["time"] - 1.0)
task["status"] = "completed"
self._remove_task(task["task_id"], s)
s.completed_tasks.append(task)
s.tasks_completed += 1
# Reward: base value × priority, bonus for early completion
deadline_bonus = max(0.0, deadline / max(time_left, 1)) * 0.5
reward = value * priority + deadline_bonus
s.cumulative_value += reward
return reward, (
f"Successfully executed '{task['name']}'. "
f"Reward: {reward:.2f} (value={value:.1f}, priority={priority:.2f})."
)
elif action_type == ActionType.DEFER:
# Low penalty; task goes to deferred queue
task["status"] = "deferred"
self._remove_task(task["task_id"], s)
self._deferred_tasks.append(task)
self._sync_state_queues()
s.resources["time"] = max(0.0, s.resources["time"] - 0.5)
defer_penalty = -0.2 * priority # higher priority = bigger penalty for deferring
return defer_penalty, (
f"Deferred '{task['name']}'. "
f"Penalty: {defer_penalty:.2f}. Will retry when resources recover."
)
elif action_type == ActionType.DELEGATE:
# Offload — no resource cost, reduced reward
task["status"] = "completed"
self._remove_task(task["task_id"], s)
s.completed_tasks.append(task)
s.tasks_completed += 1
delegate_reward = value * priority * 0.35
s.cumulative_value += delegate_reward
return delegate_reward, (
f"Delegated '{task['name']}'. "
f"Reward: {delegate_reward:.2f} (35% of full value)."
)
elif action_type == ActionType.OPTIMIZE:
# Spend a small energy/budget to reduce uncertainty
opt_energy = max(0.3, energy_cost * 0.2)
opt_budget = max(1.0, budget_cost * 0.15)
if s.resources["energy"] < opt_energy:
return -0.1, f"Cannot optimize '{task['name']}': not enough energy."
s.resources["energy"] = max(0.0, s.resources["energy"] - opt_energy)
s.resources["budget"] = max(0.0, s.resources["budget"] - opt_budget)
s.resources["time"] = max(0.0, s.resources["time"] - 0.5)
# Reduce uncertainty significantly
reduction = self._rng.uniform(0.2, 0.45)
old_unc = task["uncertainty"]
task["uncertainty"] = max(0.02, task["uncertainty"] - reduction)
return 0.1, (
f"Optimized '{task['name']}': uncertainty {old_unc:.2f} → {task['uncertainty']:.2f}. "
f"Small positive reward for risk reduction."
)
return 0.0, "Unknown action type."
def _find_task(self, task_id: int, task_list: List[Dict[str, Any]]) -> Optional[Dict[str, Any]]:
for t in task_list:
if t["task_id"] == task_id and t["status"] in ("pending", "deferred"):
return t
return None
def _remove_task(self, task_id: int, s: AetherTaskFlowState) -> None:
s.tasks = [t for t in s.tasks if t["task_id"] != task_id]
self._deferred_tasks = [t for t in self._deferred_tasks if t["task_id"] != task_id]
self._sync_state_queues()
def _sync_state_queues(self) -> None:
self._state.deferred_tasks = [dict(task) for task in self._deferred_tasks]
def _max_positive_step_reward(self) -> float:
"""Upper-bound the raw positive reward for the current difficulty profile."""
priority_max = float(self._profile["priority_range"][1])
value_max = float(self._profile["value_range"][1])
deadline_max = float(self._profile["deadline_range"][1])
return max(1.0, (value_max * priority_max) + (deadline_max * 0.5))
def _normalize_step_reward(self, raw_reward: float) -> float:
"""
Map raw action rewards into [0, 1] for OpenEnv-facing observations.
Negative rewards occupy [0.0, 0.5), zero remains a neutral midpoint in
the action-reward scale, and positive rewards occupy (0.5, 1.0].
Non-action lifecycle observations (reset/already-done) still emit their
explicit reward values directly via _build_obs.
"""
min_reward = -1.0
if raw_reward <= 0.0:
normalized = ((raw_reward - min_reward) / (0.0 - min_reward)) * 0.5
else:
max_reward = self._max_positive_step_reward()
normalized = 0.5 + 0.5 * min(raw_reward / max_reward, 1.0)
return round(max(0.0, min(1.0, normalized)), 4)
def _get_obs(self) -> Dict[str, Any]:
"""Return a compact state snapshot for manual debugging and simple UIs."""
s = self._state
self._sync_state_queues()
return {
"episode_id": s.episode_id,
"difficulty": s.difficulty,
"step_count": s.step_count,
"num_tasks": len(s.tasks),
"num_deferred_tasks": len(self._deferred_tasks),
"tasks_completed": s.tasks_completed,
"tasks_failed": s.tasks_failed,
"resources": {
"time": round(s.resources.get("time", 0.0), 2),
"energy": round(s.resources.get("energy", 0.0), 2),
"budget": round(s.resources.get("budget", 0.0), 2),
},
"system_health": round(s.system_health, 2),
"done": s.episode_done,
}
def _make_task_info(self, t: Dict[str, Any]):
from env.tasks import TaskInfo as _TI
return _TI(
task_id=t["task_id"],
name=t["name"],
priority=t["priority"],
deadline=t["deadline"],
uncertainty=t["uncertainty"],
value=t["value"],
required_energy=t["required_energy"],
required_budget=t["required_budget"],
category=t["category"],
status=t["status"],
)
def _build_obs(
self,
last_action_type: Optional[str],
last_action_task_id: Optional[int],
last_action_outcome: Optional[str],
reward: float,
done: bool,
) -> AetherTaskFlowObservation:
s = self._state
return AetherTaskFlowObservation(
done=done,
reward=reward,
metadata={
"difficulty": s.difficulty,
"episode_id": s.episode_id,
"step_count": s.step_count,
"summary": self._get_obs(),
},
tasks=list(s.tasks),
time_remaining=int(s.resources.get("time", 0)),
energy_remaining=round(s.resources.get("energy", 0.0), 2),
budget_remaining=round(s.resources.get("budget", 0.0), 2),
system_health=round(s.system_health, 4),
step_number=s.step_count,
tasks_completed=s.tasks_completed,
tasks_failed=s.tasks_failed,
cumulative_value=round(s.cumulative_value, 4),
last_action_type=last_action_type,
last_action_task_id=last_action_task_id,
last_action_outcome=last_action_outcome,
difficulty=s.difficulty,
episode_id=s.episode_id,
)
def _build_grade_result(self) -> Dict[str, Any]:
s = self._state
profile = self._profile
return {
"difficulty": s.difficulty,
"tasks_completed": s.tasks_completed,
"tasks_failed": s.tasks_failed,
"total_tasks": s.tasks_completed + s.tasks_failed + len(s.tasks) + len(self._deferred_tasks),
"remaining_time": s.resources.get("time", 0),
"remaining_energy": s.resources.get("energy", 0),
"remaining_budget": s.resources.get("budget", 0),
"initial_time": s.initial_resources.get("time", profile["initial_time"]),
"initial_energy": s.initial_resources.get("energy", profile["initial_energy"]),
"initial_budget": s.initial_resources.get("budget", profile["initial_budget"]),
"system_health": s.system_health,
"steps_used": s.step_count,
"max_steps": profile["max_steps"],
"cumulative_value": s.cumulative_value,
}
def compute_final_score(self) -> float:
"""Compute the final grade [0, 1] for the completed episode."""
result = self._build_grade_result()
score = grade(self._difficulty, result)
return max(0.0, min(1.0, score))
|