Spaces:
Sleeping
Sleeping
File size: 8,406 Bytes
dfbb493 | 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 | """
tasks.py β SpectraQual Task Definitions and Programmatic Graders
Each task runs the environment with a fixed seed and scores the agent 0.0β1.0.
Graders are deterministic and reproducible.
"""
from __future__ import annotations
import sys
import os
from typing import List
sys.path.insert(0, os.path.dirname(__file__))
from config import (
TASKS,
MEDIUM_ECONOMIC_TARGET,
HARD_ANOMALY_RATE_TARGET,
SUCCESS_SCORE_THRESHOLD,
)
from models import TaskResult
from env import SpectraQualEnv
from models import PCBAction
# ---------------------------
# TASK RUNNER
# ---------------------------
def run_task(task_id: str, actions: List[str]) -> TaskResult:
"""
Run a task with a pre-determined list of actions.
Used by graders to replay an agent's trajectory deterministically.
Args:
task_id: one of "task_easy", "task_medium", "task_hard"
actions: list of action strings, one per step
Returns:
TaskResult with all episode metrics filled in.
"""
cfg = TASKS[task_id]
env = SpectraQualEnv(task_id=task_id)
env.reset()
rewards: List[float] = []
correct = 0
total = 0
bottlenecks = 0
anomaly_total = 0
anomaly_flagged = 0
cum_raw = 0.0
for i, action_str in enumerate(actions):
if env._done:
break
# Default to SCRAP if action is out of valid range
valid = env._current_pcb and env._current_pcb.get("defect_type")
try:
result = env.step(PCBAction(action=action_str))
except Exception:
result = env.step(PCBAction(action="SCRAP"))
rewards.append(result.reward)
total += 1
if result.info.get("is_anomaly"):
anomaly_total += 1
if result.reward_components:
cum_raw += result.reward_components.total_raw
if result.info.get("is_anomaly") and result.reward_components.anomaly_bonus >= 0.8:
anomaly_flagged += 1
if env._is_correct(result.info.get("defect", ""), action_str):
correct += 1
bottlenecks = env._bottleneck_cnt
max_possible_raw = cfg["n_boards"] * 1.0 # max normalized = 1.0 per step
return TaskResult(
task_id=task_id,
total_steps=total,
rewards=rewards,
correct_decisions=correct,
total_decisions=total,
bottleneck_count=bottlenecks,
anomaly_total=anomaly_total,
anomaly_flagged=anomaly_flagged,
cumulative_raw_reward=cum_raw,
max_possible_raw=max_possible_raw,
)
# ---------------------------
# GRADER: TASK EASY
# ---------------------------
def grade_easy(result: TaskResult) -> float:
"""
Task Easy Grader.
Objective: Correctly classify all defect types. No slot pressure.
Scoring: correct_decisions / total_decisions β 0.0β1.0
Also gives partial credit for near-correct results:
- 100% correct = 1.0
- 80% correct = 0.8
- 0% correct = 0.0
"""
if result.total_decisions == 0:
return 0.0
accuracy = result.correct_decisions / result.total_decisions
# Blend accuracy with average reward for robustness
avg_reward = sum(result.rewards) / len(result.rewards) if result.rewards else 0.0
# Weight: 70% accuracy, 30% reward quality
score = 0.70 * accuracy + 0.30 * avg_reward
return round(min(max(score, 0.0), 1.0), 4)
# ---------------------------
# GRADER: TASK MEDIUM
# ---------------------------
def grade_medium(result: TaskResult) -> float:
"""
Task Medium Grader.
Objective: Triage 15 boards with 1 slot (queue pressure).
Scoring: 0.6 * economic_efficiency + 0.4 * bottleneck_avoidance
- economic_efficiency: avg normalized reward vs target
- bottleneck_avoidance: 1.0 if no bottlenecks, scales down to 0
"""
if not result.rewards:
return 0.0
avg_reward = sum(result.rewards) / len(result.rewards)
# Economic efficiency: how close to target (MEDIUM_ECONOMIC_TARGET = 0.50)
economic_score = min(avg_reward / MEDIUM_ECONOMIC_TARGET, 1.0)
# Bottleneck avoidance: 0 bottleneck = 1.0, β₯5 = 0.0
max_tolerable_bottlenecks = 5
bottleneck_score = max(0.0, 1.0 - result.bottleneck_count / max_tolerable_bottlenecks)
score = 0.60 * economic_score + 0.40 * bottleneck_score
return round(min(max(score, 0.0), 1.0), 4)
# ---------------------------
# GRADER: TASK HARD
# ---------------------------
def grade_hard(result: TaskResult) -> float:
"""
Task Hard Grader.
Objective: 20 boards, mixed anomalies, tight slots.
Scoring: 0.5 * anomaly_score + 0.3 * economic_score + 0.2 * throughput_score
- anomaly_score: anomaly_flagged / max(anomaly_total, 1), target β₯ 0.5
- economic_score: avg normalized reward
- throughput_score: boards_processed / total (penalizes WAIT spam)
"""
if not result.rewards:
return 0.0
cfg = TASKS["task_hard"]
avg_reward = sum(result.rewards) / len(result.rewards)
# Anomaly score: did the agent handle anomalous boards correctly?
if result.anomaly_total > 0:
raw_anomaly = result.anomaly_flagged / result.anomaly_total
else:
raw_anomaly = 1.0 # no anomalies β not penalized
# Scale anomaly score: meeting HARD_ANOMALY_RATE_TARGET = 1.0
anomaly_score = min(raw_anomaly / HARD_ANOMALY_RATE_TARGET, 1.0)
# Economic score
economic_score = avg_reward
# Throughput: penalize excessive WAIT actions
throughput_score = min(result.total_decisions / cfg["n_boards"], 1.0)
score = (
0.50 * anomaly_score +
0.30 * economic_score +
0.20 * throughput_score
)
return round(min(max(score, 0.0), 1.0), 4)
# ---------------------------
# GRADER DISPATCH
# ---------------------------
GRADERS = {
"task_easy": grade_easy,
"task_medium": grade_medium,
"task_hard": grade_hard,
}
def grade(task_id: str, result: TaskResult) -> float:
"""Dispatch to the correct grader for the given task_id."""
if task_id not in GRADERS:
raise ValueError(f"No grader for task_id='{task_id}'")
return GRADERS[task_id](result)
# ---------------------------
# TASK DESCRIPTIONS (for README / inference prompt)
# ---------------------------
TASK_DESCRIPTIONS = {
"task_easy": (
"Triage 10 PCBs with no factory slot pressure. "
"Focus: identify the correct action for each defect type. "
"Grader: accuracy-weighted reward (70% accuracy + 30% reward quality). "
"Expected frontier model score: β₯0.85."
),
"task_medium": (
"Triage 15 PCBs with only 1 active soldering slot. "
"Focus: manage queue pressure while maintaining economic performance. "
"Grader: 60% economic efficiency + 40% bottleneck avoidance. "
"Expected frontier model score: β₯0.65."
),
"task_hard": (
"Triage 20 PCBs with 25% anomaly rate and tight slot constraints. "
"Focus: handle extreme-cost/criticality boards safely AND maintain throughput. "
"Grader: 50% anomaly handling + 30% economic score + 20% throughput. "
"Expected frontier model score: β₯0.50."
),
}
# ---------------------------
# CLI TEST UTILITY
# ---------------------------
if __name__ == "__main__":
"""Quick sanity check: run all 3 tasks with a rule-based agent."""
from env import SpectraQualEnv, decide_action
from models import PCBAction
print("\n=== SpectraQual Task Grader Sanity Check ===\n")
for tid in ["task_easy", "task_medium", "task_hard"]:
env = SpectraQualEnv(task_id=tid)
result_obj = env.reset()
actions = []
while not result_obj.done:
obs = result_obj.observation
pcb = {
"defect_type": obs.defect_type,
"component_cost": obs.component_cost,
"criticality": obs.criticality,
}
action_str = decide_action(pcb)
actions.append(action_str)
result_obj = env.step(PCBAction(action=action_str))
task_result = run_task(tid, actions)
score = grade(tid, task_result)
print(f"[{tid}] Score: {score:.4f} | Correct: {task_result.correct_decisions}/{task_result.total_decisions} | Bottlenecks: {task_result.bottleneck_count}")
print("\n=== Done ===")
|