File size: 17,918 Bytes
77da5ce | 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 | import math
import copy
import json
import re
from core.life_state import LifeMetrics
from core.task import Task
def compute_reward(
state_before: LifeMetrics,
state_after: LifeMetrics,
resources_used: dict,
actions_taken: int,
metric_changes: dict = None,
completion: str = None,
disruption_baseline: int = None,
action_type: str = ""
) -> tuple[float, dict]:
"""
Computes the reward for a life step based on changes in LifeMetrics and resource usage.
Args:
state_before: The state at the start of the step.
state_after: The state after actions and cascades.
resources_used: Dict with keys 'time', 'money', 'energy'.
actions_taken: Integer count of intentional actions performed.
disruption_baseline: Expected number of metrics affected by an action.
Returns:
tuple[float, dict]: (final_reward, breakdown_dict)
"""
before_flat = state_before.flatten()
after_flat = state_after.flatten()
# 1. OUTCOME SCORE (Weighted average of positive deltas)
domain_weights = {
"career": 1/6,
"finances": 1/6,
"relationships": 1/6,
"physical_health": 1/6,
"mental_wellbeing": 1/6,
"time": 1/6
}
# Map sub-metrics to their domains
submetrics_per_domain = {}
for k in before_flat.keys():
domain = k.split('.')[0]
submetrics_per_domain[domain] = submetrics_per_domain.get(domain, 0) + 1
outcome_score = 0.0
for k in before_flat.keys():
domain = k.split('.')[0]
delta = after_flat[k] - before_flat[k]
if delta > 0:
# Each domain is 1/6. Each sub-metric within a domain gets its equal share of that 1/6.
# Normalize delta by 100 (max possible increase is 100).
weight = domain_weights[domain] / submetrics_per_domain[domain]
outcome_score += (delta / 100.0) * weight
# 2. CASCADE CONTAINMENT SCORE
worsened_count = sum(1 for k in before_flat.keys() if after_flat[k] < before_flat[k])
total_metrics = len(before_flat)
cascade_containment_score = 1.0 - (worsened_count / total_metrics)
# 3. RESOURCE EFFICIENCY SCORE
# Available: time 20, money 500, energy 100
m_time = resources_used.get('time', 0.0) / 20.0
m_money = resources_used.get('money', 0.0) / 500.0
m_energy = resources_used.get('energy', 0.0) / 100.0
# Normalize by total slots (3 resources)
resource_efficiency_score = 1.0 - ((m_time + m_money + m_energy) / 3.0)
resource_efficiency_score = max(0.0, min(1.0, resource_efficiency_score))
# 4. RELATIONSHIP PRESERVATION SCORE (Sigmoid applied to average delta)
rel_keys = [k for k in before_flat.keys() if k.startswith('relationships.')]
avg_rel_before = sum(before_flat[k] for k in rel_keys) / len(rel_keys)
avg_rel_after = sum(after_flat[k] for k in rel_keys) / len(rel_keys)
delta_rel = avg_rel_after - avg_rel_before
# score = 1 / (1 + exp(-delta/10))
relationship_preservation_score = 1.0 / (1.0 + math.exp(-delta_rel / 10.0))
# FINAL REWARD FORMULA
base_reward = (
(0.40 * outcome_score) +
(0.25 * cascade_containment_score) +
(0.20 * resource_efficiency_score) +
(0.15 * relationship_preservation_score)
)
# PENALTIES
penalties = 0.0
fired = []
# -0.50 if ANY metric is below 20 after the step
if any(v < 20 for v in after_flat.values()):
penalties -= 0.50
fired.append("CRITICAL_FLOOR_VIOLATION")
# -0.30 if cascade spread wider than the number of metrics the agent directly changed
# Scaled baseline from task metadata preferred over hardcoded default
if disruption_baseline is None:
disruption_baseline = len(metric_changes) if metric_changes else 2
if worsened_count > disruption_baseline:
penalties -= 0.30
fired.append("CASCADE_SPREAD_WIDER")
# -0.40 if actions_taken == 0
if actions_taken == 0:
penalties -= 0.40
fired.append("INACTION_PENALTY")
# -0.15 if relationships domain average dropped more than 20 points
if delta_rel < -20:
penalties -= 0.15
fired.append("RELATIONSHIP_COLLAPSE")
# [NEW] Plausibility Penalty
plaus = 0.0
if metric_changes:
plaus = reward_plausibility_check(metric_changes, resources_used)
if plaus < 0:
penalties += plaus
fired.append("PLAUSIBILITY_VIOLATION")
# [NEW] Format Compliance & Reasoning
comp_reward = 0.0
reasoning = ""
if completion:
comp_reward = reward_format_compliance(completion)
try:
# Simple extract reasoning from JSON if possible
import json
data = json.loads(completion)
reasoning = data.get("reasoning", "")
except:
pass
# [NEW] Reasoning Alignment (tied to action_type)
reasoning_score = reward_reasoning_coherence(reasoning, action_type=action_type)
final_reward = max(-1.0, min(1.0, base_reward + penalties))
breakdown = {
"components": {
"outcome": outcome_score,
"containment": cascade_containment_score,
"efficiency": resource_efficiency_score,
"preservation": relationship_preservation_score,
"format_compliance": comp_reward,
"plausibility": plaus,
"reasoning_alignment": reasoning_score
},
"base_reward": base_reward,
"penalties_total": penalties,
"penalties_fired": fired,
"metrics_worsened": worsened_count,
"rel_delta": delta_rel
}
return final_reward, breakdown
def compute_milestone_reward(milestones_achieved: list[str], task: Task) -> float:
if not task.milestones:
return 0.0
total_possible = sum(m.reward for m in task.milestones)
if total_possible == 0:
return 0.0
achieved = sum(m.reward for m in task.milestones if m.id in milestones_achieved)
return min(1.0, achieved / total_possible)
def compute_task_completion_reward(success_conditions_met: list[bool], task: Task) -> float:
# A task is completed if any of its target success conditions are satisfied.
# This handles tasks with multiple alternative goal-states (e.g. choice of routes).
if not success_conditions_met:
return 0.0
return 1.0 if any(success_conditions_met) else 0.0
def compute_replan_bonus(exo_events_seen: int, milestones_after_event: int) -> float:
# Scale bonus based on ability to bounce back after exogenous events
if exo_events_seen == 0:
return 0.0
return min(1.0, (milestones_after_event / exo_events_seen) * 0.5)
def compute_dead_end_penalty(routes_remaining: int) -> float:
return -0.5 if routes_remaining <= 0 else 0.0
def compute_task_reward(
state_before: LifeMetrics,
state_after: LifeMetrics,
resources_used: dict,
actions_taken: int,
milestones_achieved: list[str],
success_conditions_met: list[bool],
exo_events_seen: int,
milestones_after_event: int,
routes_remaining: int,
rollback_used: bool,
cascade_collapse: bool,
task: Task,
reasoning: str = "",
completion: str = "",
conflict_domain: str = "",
step_count: int = 0,
max_steps: int = 0,
metric_changes: dict = None,
cumulative_rel_delta: float = 0.0,
action_type: str = ""
) -> tuple[float, dict]:
# 1. Base local components (with scaled disruption baseline from task metadata)
d_baseline = len(task.mutable_world) if task and hasattr(task, 'mutable_world') else None
local_reward, local_breakdown = compute_reward(state_before, state_after, resources_used, actions_taken,
metric_changes=metric_changes, completion=completion,
disruption_baseline=d_baseline, action_type=action_type)
# 2. Orchestrator components
# Use only the raw outcome component from local_breakdown to avoid double-counting
# efficiency, containment, or preservation which are added separately below.
outcome_score_local = local_breakdown["components"].get("outcome", 0.0)
milestone_score = compute_milestone_reward(milestones_achieved, task)
completion_score = compute_task_completion_reward(success_conditions_met, task)
replan_score = compute_replan_bonus(exo_events_seen, milestones_after_event)
efficiency_score = local_breakdown["components"].get("efficiency", 0.0)
preservation_score = local_breakdown["components"].get("preservation", 0.0)
reasoning_score = reward_reasoning_coherence(reasoning, action_type=action_type)
# Check for specific failure cases
timeout_pen = reward_timeout_check(step_count, max_steps, any(success_met for success_met in success_conditions_met) if success_conditions_met else False)
dead_end_pen = compute_dead_end_penalty(routes_remaining)
# 3. Final weighting (all components are now unique/non-overlapping)
# Weights: Milestone 35%, Completion 25%, Outcome 10%, Preservation 5%, Replan 10%, Efficiency 10%, Reasoning 5%
base_reward = (
(0.35 * milestone_score) +
(0.25 * completion_score) +
(0.10 * outcome_score_local) +
(0.05 * preservation_score) +
(0.10 * replan_score) +
(0.10 * efficiency_score) +
(0.05 * reasoning_score)
)
# 4. Penalties
penalties = 0.0
fired = []
if timeout_pen < 0:
penalties += timeout_pen
fired.append("TIMEOUT")
if dead_end_pen < 0:
penalties += dead_end_pen
fired.append("DEAD_END")
if rollback_used:
penalties += -0.1
fired.append("ROLLBACK_USED")
if cascade_collapse:
penalties += -0.3
fired.append("CASCADE_COLLAPSE")
# Direct inaction penalty — not diluted by the 0.05 local weight
if actions_taken == 0:
penalties += -0.20
fired.append("TASK_INACTION_PENALTY")
# Cumulative relationship erosion across the episode
if cumulative_rel_delta < -20:
penalties += -0.15
fired.append("CUMULATIVE_RELATIONSHIP_EROSION")
final_reward = max(-1.0, min(1.0, base_reward + penalties))
breakdown = {
"components": {
"local_metric_delta": outcome_score_local,
"milestone": milestone_score,
"completion": completion_score,
"replan": replan_score,
"efficiency": efficiency_score,
"reasoning": reasoning_score,
"format_compliance": local_breakdown["components"].get("format_compliance", 0.0),
"plausibility": local_breakdown["components"].get("plausibility", 0.0),
"timeout_penalty": timeout_pen
},
"base_reward": base_reward,
"penalties_total": penalties,
"penalties_fired": fired,
"local_breakdown": local_breakdown
}
return final_reward, breakdown
def reward_format_compliance(completion: str) -> float:
"""
Scores the completion based on its format (JSON validity and required fields).
Returns:
+1.0: Valid JSON with all required fields:
action_type, target_domain, metric_changes, resource_cost, reasoning
+0.5: Any parseable JSON (including partial/incomplete dicts)
-0.5: Invalid JSON / unparseable
-1.0: Empty strings or refusal content
"""
if not completion or len(completion.strip()) < 10:
return -1.0
# Potential refusal indicators
if any(x in completion.lower() for x in ["i cannot", "i'm sorry", "as an ai"]):
return -1.0
# Extract JSON content from markdown code blocks if present
json_str = completion.strip()
if "```json" in json_str:
json_str = json_str.split("```json")[-1].split("```")[0].strip()
elif "```" in json_str:
json_str = json_str.split("```")[-1].split("```")[0].strip()
try:
data = json.loads(json_str)
required = ["action_type", "target_domain", "metric_changes", "resource_cost", "reasoning"]
if isinstance(data, dict) and all(k in data and data.get(k) is not None for k in required):
return 1.0
return 0.5
except json.JSONDecodeError:
# Final attempt: try to find anything between { and }
match = re.search(r'\{.*\}', json_str, re.DOTALL)
if match:
try:
data = json.loads(match.group(0))
required = ["action_type", "target_domain", "metric_changes", "resource_cost", "reasoning"]
if isinstance(data, dict) and all(k in data and data.get(k) is not None for k in required):
return 1.0
return 0.5
except:
pass
return -0.5
def reward_plausibility_check(metric_changes: dict, resource_cost: dict) -> float:
"""
Anti-gaming check. Prevents the model from claiming massive metric changes while spending 0 resources.
Resource cost is normalized to comparable units (time/20h, money/$500, energy/100pts).
"""
total_delta = sum(abs(v) for v in metric_changes.values())
# Zero-cost shortcut: any non-trivial claim with no cost at all is implausible
# Also handles empty resource_cost.
if not resource_cost or all(v == 0 for v in resource_cost.values()):
if total_delta > 3.0:
return -0.30
return 0.0
# Normalize each resource dimension to [0,1] before summing
norm_time = resource_cost.get('time', 0.0) / 20.0
norm_money = resource_cost.get('money', 0.0) / 500.0
norm_energy = resource_cost.get('energy', 0.0) / 100.0
total_cost = norm_time + norm_money + norm_energy
ratio = total_delta / max(0.01, total_cost)
if ratio > 150:
return -0.30 # Claiming massive change for virtually free
if ratio > 80:
return -0.10 # Highly suspicious efficiency
return 0.0 # Plausible ratio
def reward_timeout_check(step_count: int, max_steps: int, done: bool) -> float:
"""
Penalizes episodes that end by reaching the step limit without being resolved.
"""
if step_count >= max_steps and not done:
return -0.20
return 0.0
def reward_reasoning_coherence(reasoning: str, action_type: str = "") -> float:
"""
Harden verification of logical consistency. Requires both length and
alignment with the chosen action to prevent word-stuffing.
"""
if not reasoning or len(reasoning.strip()) < 20:
return -0.20 # Severe penalty for lack of effort
reasoning_lower = reasoning.lower()
score = 0.0
# 1. Structural Logic Check
# Reward use of logical connectors rather than just list of facts
connectors = ["because", "since", "therefore", "due to", "resulting in", "consequently"]
if any(c in reasoning_lower for c in connectors):
score += 0.05
# 2. Action Alignment (Non-Gammable Anti-Hacking)
# The reasoning MUST logically justify the chosen category.
action_keywords = {
"spend": ["cost", "price", "expensive", "money", "budget", "finance"],
"rest": ["energy", "sleep", "exhaustion", "recharge", "break"],
"communicate": ["talk", "discuss", "speak", "message", "call", "explain"],
"delegate": ["hand off", "assign", "help", "junior", "colleague"],
"negotiate": ["bargain", "trade", "deal", "terms"],
"deprioritize": ["later", "postpone", "unimportant", "drop"],
"reschedule": ["reschedule", "delay", "postpone", "move", "time", "calendar", "slot"],
"execute": ["route", "plan", "action", "implement", "complete", "resolve", "execute"],
}
if action_type and action_type in action_keywords:
match = any(kw in reasoning_lower for kw in action_keywords[action_type])
if match:
score += 0.10
else:
score -= 0.20
return max(-0.30, min(0.30, score))
def main():
# Scenario setup
print("--- TESTING REWARD SYSTEM ---")
# 1. PERFECT ACTION: All metrics improve by 10 points
state_start = LifeMetrics() # Defaults at 70
state_perfect = copy.deepcopy(state_start)
for k in state_perfect.flatten().keys():
domain, sub = k.split('.')
current = getattr(getattr(state_perfect, domain), sub)
setattr(getattr(state_perfect, domain), sub, current + 10)
res_perfect = {"time": 2, "money": 50, "energy": 10}
reward_p, break_p = compute_reward(state_start, state_perfect, res_perfect, actions_taken=5)
print("\n[SCENARIO 1: PERFECT ACTION]")
print(f"Reward: {reward_p:.4f}")
print(f"Breakdown: {break_p}")
# 2. BAD ACTION: Relationships tank by 30 points, everything else stays same
state_bad = copy.deepcopy(state_start)
for k in state_bad.flatten().keys():
if k.startswith('relationships.'):
domain, sub = k.split('.')
current = getattr(getattr(state_bad, domain), sub)
setattr(getattr(state_bad, domain), sub, current - 30)
res_bad = {"time": 10, "money": 300, "energy": 80}
reward_b, break_b = compute_reward(state_start, state_bad, res_bad, actions_taken=1)
print("\n[SCENARIO 2: BAD ACTION (Relationships Tank)]")
print(f"Reward: {reward_b:.4f}")
print(f"Breakdown: {break_b}")
# 3. INACTION: Nothing changes
state_nothing = copy.deepcopy(state_start)
res_none = {}
reward_n, break_n = compute_reward(state_start, state_nothing, res_none, actions_taken=0)
print("\n[SCENARIO 3: INACTION]")
print(f"Reward: {reward_n:.4f}")
print(f"Breakdown: {break_n}")
if __name__ == "__main__":
main()
|