Spaces:
Running
Running
File size: 11,188 Bytes
40ee6b4 |
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 |
"""
MCTS Policies Module - Selection, rollout, and evaluation policies.
Provides:
- UCB1 with configurable exploration weight
- Rollout heuristics (random, greedy, hybrid)
- Action selection policies (max visits, max value, robust child)
- Progressive widening parameters
"""
from __future__ import annotations
import math
from abc import ABC, abstractmethod
from collections.abc import Awaitable, Callable
from enum import Enum
from typing import TYPE_CHECKING
import numpy as np
if TYPE_CHECKING:
from .core import MCTSState
def ucb1(
value_sum: float,
visits: int,
parent_visits: int,
c: float = 1.414,
) -> float:
"""
Upper Confidence Bound 1 (UCB1) formula for tree selection.
Formula: Q(s,a) + c * sqrt(N(s)) / sqrt(N(s,a))
Args:
value_sum: Total accumulated value for the node
visits: Number of visits to the node
parent_visits: Number of visits to the parent node
c: Exploration weight constant (default sqrt(2))
Returns:
UCB1 score for node selection
"""
if visits == 0:
return float("inf")
exploitation = value_sum / visits
exploration = c * ((parent_visits) ** 0.5 / (visits) ** 0.5)
return exploitation + exploration
def ucb1_tuned(
value_sum: float,
value_squared_sum: float,
visits: int,
parent_visits: int,
c: float = 1.0,
) -> float:
"""
UCB1-Tuned variant with variance estimate.
Provides tighter bounds by considering value variance.
Args:
value_sum: Total accumulated value
value_squared_sum: Sum of squared values (for variance)
visits: Number of visits
parent_visits: Parent visit count
c: Exploration constant
Returns:
UCB1-Tuned score
"""
if visits == 0:
return float("inf")
mean_value = value_sum / visits
variance = value_squared_sum / visits - mean_value**2
variance = max(0, variance) # Ensure non-negative
# Variance bound term
ln_parent = math.log(parent_visits)
variance_bound = variance + math.sqrt(2 * ln_parent / visits)
min_bound = min(0.25, variance_bound)
exploitation = mean_value
exploration = c * math.sqrt(ln_parent / visits * min_bound)
return exploitation + exploration
class SelectionPolicy(Enum):
"""Policy for selecting the final action after MCTS search."""
MAX_VISITS = "max_visits"
"""Select action with most visits (most robust)."""
MAX_VALUE = "max_value"
"""Select action with highest average value (greedy)."""
ROBUST_CHILD = "robust_child"
"""Select action balancing visits and value."""
SECURE_CHILD = "secure_child"
"""Select action with lowest lower confidence bound."""
class RolloutPolicy(ABC):
"""Abstract base class for rollout/simulation policies."""
@abstractmethod
async def evaluate(
self,
state: MCTSState,
rng: np.random.Generator,
max_depth: int = 10,
) -> float:
"""
Evaluate a state through rollout simulation.
Args:
state: State to evaluate
rng: Seeded random number generator
max_depth: Maximum rollout depth
Returns:
Estimated value in [0, 1] range
"""
pass
class RandomRolloutPolicy(RolloutPolicy):
"""Random rollout policy - uniform random evaluation."""
def __init__(self, base_value: float = 0.5, noise_scale: float = 0.3):
"""
Initialize random rollout policy.
Args:
base_value: Base value for evaluations
noise_scale: Scale of random noise
"""
self.base_value = base_value
self.noise_scale = noise_scale
async def evaluate(
self,
_state: MCTSState,
rng: np.random.Generator,
_max_depth: int = 10,
) -> float:
"""Generate random evaluation with noise."""
noise = rng.uniform(-self.noise_scale, self.noise_scale)
value = self.base_value + noise
return max(0.0, min(1.0, value))
class GreedyRolloutPolicy(RolloutPolicy):
"""Greedy rollout policy using domain heuristics."""
def __init__(
self,
heuristic_fn: Callable[[MCTSState], float],
noise_scale: float = 0.05,
):
"""
Initialize greedy rollout policy.
Args:
heuristic_fn: Function to evaluate state heuristically
noise_scale: Small noise for tie-breaking
"""
self.heuristic_fn = heuristic_fn
self.noise_scale = noise_scale
async def evaluate(
self,
state: MCTSState,
rng: np.random.Generator,
_max_depth: int = 10,
) -> float:
"""Evaluate using heuristic with small noise."""
base_value = self.heuristic_fn(state)
noise = rng.uniform(-self.noise_scale, self.noise_scale)
value = base_value + noise
return max(0.0, min(1.0, value))
class HybridRolloutPolicy(RolloutPolicy):
"""Hybrid policy combining random and heuristic evaluation."""
def __init__(
self,
heuristic_fn: Callable[[MCTSState], float] | None = None,
heuristic_weight: float = 0.7,
random_weight: float = 0.3,
base_random_value: float = 0.5,
noise_scale: float = 0.2,
):
"""
Initialize hybrid rollout policy.
Args:
heuristic_fn: Optional heuristic evaluation function
heuristic_weight: Weight for heuristic component
random_weight: Weight for random component
base_random_value: Base value for random component
noise_scale: Noise scale for random component
"""
self.heuristic_fn = heuristic_fn
self.heuristic_weight = heuristic_weight
self.random_weight = random_weight
self.base_random_value = base_random_value
self.noise_scale = noise_scale
# Normalize weights
total_weight = heuristic_weight + random_weight
if total_weight > 0:
self.heuristic_weight /= total_weight
self.random_weight /= total_weight
async def evaluate(
self,
state: MCTSState,
rng: np.random.Generator,
_max_depth: int = 10,
) -> float:
"""Combine heuristic and random evaluation."""
# Random component
random_noise = rng.uniform(-self.noise_scale, self.noise_scale)
random_value = self.base_random_value + random_noise
# Heuristic component
heuristic_value = self.heuristic_fn(state) if self.heuristic_fn is not None else self.base_random_value
# Combine
value = self.heuristic_weight * heuristic_value + self.random_weight * random_value
return max(0.0, min(1.0, value))
class LLMRolloutPolicy(RolloutPolicy):
"""Rollout policy that uses an LLM for state evaluation."""
def __init__(
self,
evaluate_fn: Callable[[MCTSState], Awaitable[float]],
cache_results: bool = True,
):
"""
Initialize LLM rollout policy.
Args:
evaluate_fn: Async function to evaluate state with LLM
cache_results: Whether to cache evaluation results
"""
self.evaluate_fn = evaluate_fn
self.cache_results = cache_results
self._cache: dict = {}
async def evaluate(
self,
state: MCTSState,
_rng: np.random.Generator,
_max_depth: int = 10,
) -> float:
"""Evaluate state using LLM."""
state_key = state.to_hash_key()
if self.cache_results and state_key in self._cache:
return self._cache[state_key]
value = await self.evaluate_fn(state)
value = max(0.0, min(1.0, value))
if self.cache_results:
self._cache[state_key] = value
return value
class ProgressiveWideningConfig:
"""Configuration for progressive widening in MCTS."""
def __init__(
self,
k: float = 1.0,
alpha: float = 0.5,
):
"""
Configure progressive widening parameters.
Progressive widening expands when: visits > k * num_children^alpha
Args:
k: Coefficient controlling expansion threshold
alpha: Exponent controlling growth rate
Common configurations:
- k=1.0, alpha=0.5: Moderate widening (default)
- k=2.0, alpha=0.5: Conservative (fewer expansions)
- k=0.5, alpha=0.5: Aggressive (more expansions)
- k=1.0, alpha=0.3: Very aggressive
- k=1.0, alpha=0.7: Very conservative
"""
if k <= 0:
raise ValueError("k must be positive")
if not 0 < alpha < 1:
raise ValueError("alpha must be in (0, 1)")
self.k = k
self.alpha = alpha
def should_expand(self, visits: int, num_children: int) -> bool:
"""
Check if expansion should occur.
Args:
visits: Number of visits to node
num_children: Current number of children
Returns:
True if should expand, False otherwise
"""
threshold = self.k * (num_children**self.alpha)
return visits > threshold
def min_visits_for_expansion(self, num_children: int) -> int:
"""
Calculate minimum visits needed to expand to next child.
Args:
num_children: Current number of children
Returns:
Minimum visit count for expansion
"""
threshold = self.k * (num_children**self.alpha)
return int(math.ceil(threshold))
def __repr__(self) -> str:
return f"ProgressiveWideningConfig(k={self.k}, alpha={self.alpha})"
def compute_action_probabilities(
children_stats: list[dict],
temperature: float = 1.0,
) -> list[float]:
"""
Compute action probabilities from visit counts using softmax.
Args:
children_stats: List of dicts with 'visits' key
temperature: Temperature parameter (lower = more deterministic)
Returns:
List of probabilities for each action
"""
if not children_stats:
return []
visits = np.array([c["visits"] for c in children_stats], dtype=float)
if temperature == 0:
# Deterministic: assign 1.0 to max, 0 to others
probs = np.zeros_like(visits)
probs[np.argmax(visits)] = 1.0
return probs.tolist()
# Apply temperature
scaled_visits = visits ** (1.0 / temperature)
probs = scaled_visits / scaled_visits.sum()
return probs.tolist()
def select_action_stochastic(
children_stats: list[dict],
rng: np.random.Generator,
temperature: float = 1.0,
) -> int:
"""
Stochastically select action based on visit counts.
Args:
children_stats: List of child statistics
rng: Random number generator
temperature: Temperature for softmax
Returns:
Index of selected action
"""
probs = compute_action_probabilities(children_stats, temperature)
if not probs:
raise ValueError("No actions to select from")
return rng.choice(len(probs), p=probs)
|