File size: 11,927 Bytes
aec0295 | 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 | """
Reward computation and normalization for the trading environment.
All rewards and grades are normalized to [0, 1].
"""
import numpy as np
from typing import Dict
import json
import re
# Default reward component weights
DEFAULT_WEIGHTS = {
"profit": 1.0,
"drawdown": 0.5,
"volatility": 0.3,
"sharpe": 0.5,
"overtrading": 0.1,
"hold_penalty": 0.01,
"directional_bonus": 0.3,
}
# Normalization: tanh scale factor (higher = sharper gradient near zero)
DEFAULT_NORM_SCALE = 5.0
def compute_raw_reward(
profit: float,
drawdown: float,
volatility: float,
sharpe: float,
trade_count: int,
weights: Dict[str, float] | None = None,
direction: int = 0,
price_trend: float = 0.0,
) -> float:
"""
Compute the raw (un-normalized) reward signal.
The profit signal is amplified (Γ1000) so single-step PnL fractions
produce meaningful gradient. A small hold-penalty discourages the
model from always choosing direction=0, and a directional bonus
rewards matching the market trend.
Args:
profit: Change in portfolio value (as fraction of initial).
drawdown: Current max drawdown [0, 1].
volatility: Return standard deviation.
sharpe: Sharpe ratio of returns.
trade_count: Number of trades executed this step.
weights: Component weights (uses defaults if None).
direction: Action direction (0=Hold, 1=Buy, 2=Sell).
price_trend: Signed price change fraction for the step.
Returns:
Raw reward (float, unbounded).
"""
w = weights or DEFAULT_WEIGHTS
# Amplify per-step profit so it's not buried in noise
profit_signal = w["profit"] * profit * 1000.0
# Penalties
dd_penalty = w["drawdown"] * drawdown
vol_penalty = w["volatility"] * volatility
overtrade_penalty = w["overtrading"] * (trade_count / 10.0)
# Bonuses
sharpe_bonus = w["sharpe"] * np.tanh(sharpe)
# Hold penalty: small cost for doing nothing
hold_pen = w.get("hold_penalty", 0.01) if direction == 0 else 0.0
# Directional correctness: reward matching the trend
dir_bonus = 0.0
w_dir = w.get("directional_bonus", 0.3)
if direction == 1 and price_trend > 0: # Bought into uptrend
dir_bonus = w_dir * min(abs(price_trend) * 100, 1.0)
elif direction == 2 and price_trend < 0: # Sold into downtrend
dir_bonus = w_dir * min(abs(price_trend) * 100, 1.0)
elif direction != 0: # Wrong direction
dir_bonus = -w_dir * 0.5
reward = (
profit_signal
- dd_penalty
- vol_penalty
+ sharpe_bonus
- overtrade_penalty
- hold_pen
+ dir_bonus
)
return float(reward)
def normalize_reward(
raw: float,
scale: float | None = None,
) -> float:
"""
Normalize reward to [-1, 1] using tanh scaling.
This preserves the sign (positive = good, negative = bad) and
provides smooth gradient everywhere, unlike the old min-max clip
which collapsed everything to ~0.5.
"""
s = float(scale if scale is not None else DEFAULT_NORM_SCALE)
return float(np.tanh(raw / s))
def compute_grade(metrics: Dict[str, float]) -> float:
"""
Compute the final evaluation grade [0, 1].
grade = 0.4 * normalized_profit
+ 0.3 * normalized_sharpe
+ 0.2 * (1 - normalized_drawdown)
+ 0.1 * consistency
All input metrics must already be in [0, 1].
"""
profit = np.clip(metrics.get("profit", 0.0), 0.0, 1.0)
sharpe = np.clip(metrics.get("sharpe", 0.0), 0.0, 1.0)
drawdown = np.clip(metrics.get("drawdown", 0.0), 0.0, 1.0)
consistency = np.clip(metrics.get("consistency", 0.0), 0.0, 1.0)
grade = (
0.4 * profit
+ 0.3 * sharpe
+ 0.2 * (1.0 - drawdown)
+ 0.1 * consistency
)
return float(np.clip(grade, 0.0, 1.0))
def _extract_json_action(completion: str):
match = re.search(r"<action>\s*({.*?})\s*</action>", completion, re.DOTALL)
if not match:
return None
return json.loads(match.group(1))
def _extract_prompt_state(prompt: str):
json_match = re.search(r'"state"\s*:\s*\[(.*?)\]', prompt, re.DOTALL)
if json_match:
return [float(x.strip()) for x in json_match.group(1).split(",") if x.strip()]
plain_match = re.search(r"State:\s*\[(.*?)\]", prompt, re.DOTALL)
if plain_match:
return [float(x.strip()) for x in plain_match.group(1).split(",") if x.strip()]
return None
def _extract_signal_value(prompt: str, key: str):
json_match = re.search(rf'"{key}"\s*:\s*(-?[\d\.]+)', prompt)
if json_match:
return float(json_match.group(1))
plain_match = re.search(rf"{key}\s*[:=]\s*(-?[\d\.]+)", prompt)
if plain_match:
return float(plain_match.group(1))
return None
# ββββββββββββββββββββββββββββββββββββββββββββββ
# GRPO Verifier Functions (Expert Optimized)
# ββββββββββββββββββββββββββββββββββββββββββββββ
def format_reward_func(prompts, completions, **kwargs) -> list[float]:
"""Strict format and reasoning length check."""
rewards = []
for completion in completions:
try:
if "<thought>" not in completion or "</thought>" not in completion or "<action>" not in completion or "</action>" not in completion:
rewards.append(0.0)
continue
thought = completion.split("<thought>")[1].split("</thought>")[0].strip()
if len(thought) < 150:
rewards.append(0.2)
continue
if _extract_json_action(completion) is not None:
rewards.append(1.0)
else:
rewards.append(0.4)
except Exception:
rewards.append(0.0)
return rewards
def alignment_reward_func(prompts, completions, **kwargs) -> list[float]:
"""
Ensures the <thought> matches the signals in the <prompt>.
This is the 'Anti-Hallucination' reward.
"""
rewards = []
for prompt, completion in zip(prompts, completions):
try:
ta_signal = _extract_signal_value(prompt, "ta")
is_bullish = ta_signal is not None and ta_signal > 0.2
is_bearish = ta_signal is not None and ta_signal < -0.2
thought = completion.split("<thought>")[1].split("</thought>")[0].lower()
score = 0.5 # Baseline
if is_bullish and ("bullish" in thought or "upward" in thought or "buy" in thought):
score += 0.5
elif is_bearish and ("bearish" in thought or "downward" in thought or "sell" in thought):
score += 0.5
rewards.append(score)
except Exception:
rewards.append(0.0)
return rewards
def risk_reward_func(prompts, completions, **kwargs) -> list[float]:
"""Safety Constraint: Position limits and Stop-Loss presence."""
rewards = []
for prompt, completion in zip(prompts, completions):
try:
limit = _extract_signal_value(prompt, "position_limit")
if limit is None:
limit = _extract_signal_value(prompt, "risk")
if limit is None:
limit = 1.0
data = _extract_json_action(completion)
if data is not None:
size = float(data.get("size", 0.0))
# Reward 1: Under limit
score = 0.7 if size <= limit else 0.0
# Reward 2: Logic check (Mentioning 'risk' or 'limit' in thoughts)
thought = completion.split("<thought>")[1].split("</thought>")[0].lower()
if "risk" in thought or "limit" in thought or "constraint" in thought:
score += 0.3
rewards.append(score)
else:
rewards.append(0.0)
except Exception:
rewards.append(0.0)
return rewards
def profit_reward_func(prompts, completions, **kwargs) -> list[float]:
"""
Simulated PnL: Checks if the action (direction) matches the actual
future price trend provided in the hidden 'scenario_result' metadata.
"""
rewards = []
for prompt, completion in zip(prompts, completions):
try:
data = _extract_json_action(completion)
if data is None:
rewards.append(0.0)
continue
direction = int(data.get("direction", 0))
prices = _extract_prompt_state(prompt)
if not prices or len(prices) < 2:
rewards.append(0.0)
continue
is_up_trend = prices[-1] > prices[0]
if direction == 1 and is_up_trend: # Buy in uptrend
rewards.append(1.0)
elif direction == 2 and not is_up_trend: # Sell in downtrend
rewards.append(1.0)
elif direction == 0: # Neutral
rewards.append(0.5)
else: # Wrong direction
rewards.append(0.0)
except Exception:
rewards.append(0.0)
return rewards
def governance_reward_func(prompts, completions, **kwargs) -> list[float]:
"""Self-regulation verifier: rewards actions that would pass governance
without intervention.
An agent that **self-regulates** (proposes compliant sizes, references
risk constraints in its reasoning) scores higher than one that blindly
maximises size and forces the environment to clamp it.
Scoring rubric (0-1):
+0.40 Action has valid JSON with size β€ governance limit.
+0.20 Size uses β€ 80 % of limit (conservative, professional).
+0.20 <thought> explicitly references governance keywords
(risk, limit, constraint, compliance, conservative).
+0.20 Direction is non-zero (agent is actively trading, not idle).
-0.50 Size EXCEEDS governance limit (would trigger intervention).
"""
rewards = []
for prompt, completion in zip(prompts, completions):
try:
data = _extract_json_action(completion)
if data is None:
rewards.append(0.0)
continue
size = float(data.get("size", 0.0))
direction = int(data.get("direction", 0))
limit = _extract_signal_value(prompt, "position_limit")
if limit is None:
limit = 1.0
score = 0.0
# Core compliance: within limit
if size <= limit:
score += 0.40
# Conservative bonus: using β€ 80 % of limit
if 0 < size <= limit * 0.8:
score += 0.20
else:
# Governance would intervene β penalise
score -= 0.50
# Reasoning quality: does the thought show awareness?
try:
thought = completion.split("<thought>")[1].split("</thought>")[0].lower()
governance_keywords = ["risk", "limit", "constraint", "compliance",
"conservative", "governance", "restrict",
"drawdown", "cap", "position limit"]
if any(kw in thought for kw in governance_keywords):
score += 0.20
except (IndexError, AttributeError):
pass
# Activity bonus: non-hold action
if direction != 0:
score += 0.20
rewards.append(float(np.clip(score, 0.0, 1.0)))
except Exception:
rewards.append(0.0)
return rewards
|