Spaces:
Sleeping
Sleeping
File size: 11,968 Bytes
3040767 | 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 | """Adaptive difficulty management for the HONEST environment.
The rolling accuracy window looks at records in ``state.episode_history``
that have a ``"domain"`` key and a ``"correct"`` key. Those records are
owned and written **exclusively by** ``HonestEnvironment.step()`` β this
module is a pure analyser and mutates only ``state.domain_difficulties``.
``update_difficulty`` is now a **pure side-effect-free function** w.r.t.
history: it reads history, optionally updates the difficulty scalar, and
returns ``(new_difficulty, changed)``. The caller (environment) is
responsible for flagging the relevant history record with
``"difficulty_changed": True`` if it wishes to track transitions.
"""
import random
from collections import deque
from dataclasses import dataclass, field
from typing import Dict, List, Optional, Tuple
from models.models import HonestState
# ---------------------------------------------------------------------------
# Constants
# ---------------------------------------------------------------------------
WINDOW: int = 20 # rolling accuracy window (episodes per domain)
HIGH_THRESHOLD: float = 0.70 # accuracy above this β increase difficulty
LOW_THRESHOLD: float = 0.30 # accuracy below this β decrease difficulty
MIN_DIFFICULTY: int = 1
MAX_DIFFICULTY: int = 5
HYSTERESIS_EPISODES: int = 10 # min episodes between consecutive changes
# ---------------------------------------------------------------------------
# Adaptive sampling distribution: static floor + triangular overlay
# ---------------------------------------------------------------------------
# Always-on weight per difficulty 1..5 β protects against catastrophic
# forgetting of easy-problem competence as the curriculum advances.
STATIC_FLOOR: List[float] = [0.20, 0.15, 0.10, 0.05, 0.00] # sums to 0.50
# Remaining weight (0.50) is distributed by a triangular kernel around
# the controller's current target_difficulty.
ADAPTIVE_BUDGET: float = 0.50
def triangular_overlay(target: int, total_weight: float = ADAPTIVE_BUDGET) -> List[float]:
"""Triangular distribution centered at ``target``, summing to ``total_weight``.
Difficulties are 1-indexed; returns a 5-element list.
Kernel: ``max(0, 3 - |target - d|)`` over d in [1..5], then renormalised
to ``total_weight``. At the edges (target=1 or target=5) the kernel is
clipped, so less mass lands on phantom out-of-range difficulties β but
the surviving mass is still renormalised so the overlay always sums to
``total_weight``.
"""
raw = [max(0, 3 - abs(target - d)) for d in range(1, 6)]
s = sum(raw)
if s == 0:
return [0.0] * 5
return [r * total_weight / s for r in raw]
def compute_distribution(target_difficulty: int) -> List[float]:
"""Static floor + adaptive overlay. Returns weights for difficulties 1..5.
The result is renormalised to sum to exactly 1.0 to absorb floating-point
drift, so callers can pass it directly to ``random.choices`` weights.
"""
overlay = triangular_overlay(target_difficulty)
distribution = [STATIC_FLOOR[i] + overlay[i] for i in range(5)]
total = sum(distribution)
return [d / total for d in distribution]
# ---------------------------------------------------------------------------
# DomainState + DifficultyController
# ---------------------------------------------------------------------------
@dataclass
class DomainState:
"""Per-domain state held by ``DifficultyController``."""
target_difficulty: int = 1
rolling_window: deque = field(default_factory=lambda: deque(maxlen=20))
episodes_since_last_update: int = 0
class DifficultyController:
"""Adaptive difficulty controller with a static floor.
Per-domain state: rolling 20-episode accuracy window, a target difficulty
scalar, and a cooldown counter. Hysteresis thresholds 75 / 25, cooldown
of 10 outcomes per domain.
Lifetime: one instance per ``HonestEnvironment`` (or one per training
process for the local-rollout path). The controller persists across
episode boundaries β its state is *not* reset by ``env.reset()``.
"""
UPDATE_THRESHOLD_UP: float = 0.75
UPDATE_THRESHOLD_DOWN: float = 0.25
COOLDOWN_EPISODES: int = 10
WINDOW_SIZE: int = 20
DIFFICULTY_MIN: int = MIN_DIFFICULTY
DIFFICULTY_MAX: int = MAX_DIFFICULTY
def __init__(self, domains: List[str], initial_target: int = 1) -> None:
self.domains = list(domains)
self.state: Dict[str, DomainState] = {
d: DomainState(target_difficulty=initial_target) for d in self.domains
}
# --- sampling ----------------------------------------------------------
def sample_difficulty(
self,
domain: str,
rng: Optional[random.Random] = None,
) -> int:
"""Sample a difficulty 1..5 for ``domain`` using the current distribution."""
target = self.state[domain].target_difficulty
weights = compute_distribution(target)
chooser = rng if rng is not None else random
return chooser.choices([1, 2, 3, 4, 5], weights=weights, k=1)[0]
# --- outcome tracking --------------------------------------------------
def record_outcome(self, domain: str, correct: bool) -> Tuple[int, bool]:
"""Record an episode outcome.
Returns ``(new_target_difficulty, did_update)``. Should be called
only with True/False β abstain / malformed (correct=None) episodes
must NOT enter the rolling window.
"""
s = self.state[domain]
s.rolling_window.append(1 if correct else 0)
s.episodes_since_last_update += 1
did_update = False
if (
s.episodes_since_last_update >= self.COOLDOWN_EPISODES
and len(s.rolling_window) >= self.WINDOW_SIZE
):
accuracy = sum(s.rolling_window) / len(s.rolling_window)
if (
accuracy >= self.UPDATE_THRESHOLD_UP
and s.target_difficulty < self.DIFFICULTY_MAX
):
s.target_difficulty += 1
did_update = True
elif (
accuracy <= self.UPDATE_THRESHOLD_DOWN
and s.target_difficulty > self.DIFFICULTY_MIN
):
s.target_difficulty -= 1
did_update = True
if did_update:
# Reset the cooldown but keep the rolling window β the new
# target's accuracy estimate phases in as fresh outcomes flow.
s.episodes_since_last_update = 0
return s.target_difficulty, did_update
# --- introspection -----------------------------------------------------
def get_distribution(self, domain: str) -> List[float]:
return compute_distribution(self.state[domain].target_difficulty)
def get_target(self, domain: str) -> int:
return self.state[domain].target_difficulty
def get_rolling_accuracy(self, domain: str) -> Optional[float]:
s = self.state[domain]
if len(s.rolling_window) == 0:
return None
return sum(s.rolling_window) / len(s.rolling_window)
def snapshot(self) -> Dict[str, dict]:
"""Return a JSON-serialisable snapshot for logging / debugging."""
return {
d: {
"target_difficulty": s.target_difficulty,
"rolling_accuracy": self.get_rolling_accuracy(d),
"episodes_since_update": s.episodes_since_last_update,
"window_full": len(s.rolling_window) == self.WINDOW_SIZE,
"window_size": len(s.rolling_window),
"distribution": self.get_distribution(d),
}
for d, s in self.state.items()
}
# ---------------------------------------------------------------------------
# Internal helpers
# ---------------------------------------------------------------------------
def _domain_records(state: HonestState, domain: str) -> list[dict]:
"""Return the last WINDOW episode records for *domain* from history.
Only records that contain both 'domain' and 'correct' keys are
considered (i.e. the rich records written by the environment, not
any stale auxiliary records).
"""
records = [
r for r in state.episode_history
if r.get("domain") == domain and "correct" in r
]
return records[-WINDOW:]
def _last_change_episode(state: HonestState, domain: str) -> int:
"""Return the global episode index of the most recent difficulty change
for *domain*.
Scans ``episode_history`` backwards for a record flagged with
``"difficulty_changed": True`` for the given domain.
Returns 0 if no change has ever occurred (safe to change from episode 0).
"""
for idx in range(len(state.episode_history) - 1, -1, -1):
r = state.episode_history[idx]
if r.get("domain") == domain and r.get("difficulty_changed"):
return idx
return 0
# ---------------------------------------------------------------------------
# Public API
# ---------------------------------------------------------------------------
def get_rolling_accuracy(state: HonestState, domain: str) -> float:
"""Return the rolling accuracy (0.0β1.0) for *domain* over the last
WINDOW answered episodes.
Episodes where ``correct`` is ``None`` (abstain / malformed) are treated
as incorrect. Returns 0.5 (neutral) when there are no episodes yet.
"""
records = _domain_records(state, domain)
if not records:
return 0.5 # neutral default β no change triggered
correct_count = sum(1 for r in records if r.get("correct") is True)
return correct_count / len(records)
def update_difficulty(
state: HonestState,
last_correctness: Optional[bool],
domain: Optional[str] = None,
) -> Tuple[int, bool]:
"""Evaluate rolling accuracy and adjust ``state.domain_difficulties``
in-place if a threshold is crossed.
Parameters
----------
state:
The current environment state (mutated in-place for the difficulty
scalar only β history is **not** touched here).
last_correctness:
Whether the most recent answer was correct. ``None`` for
abstain / malformed answers (counted as incorrect).
domain:
Override for the active domain. Defaults to ``state.current_domain``.
Returns
-------
(new_difficulty, changed) where ``changed`` is True iff the difficulty
scalar was actually modified this call. The caller can use this to
stamp ``"difficulty_changed": True`` onto its own history record.
"""
if domain is None:
domain = state.current_domain
current_difficulty = state.domain_difficulties.get(domain, 1)
# We need to know how many history entries exist *before* the caller
# appends the current step. The caller must append its rich record
# *before* calling this function so rolling accuracy includes the
# current result.
global_episode_idx = len(state.episode_history) - 1 # 0-indexed last item
# --- compute rolling accuracy (includes the just-appended record) ---
accuracy = get_rolling_accuracy(state, domain)
# --- hysteresis guard ---
last_change = _last_change_episode(state, domain)
episodes_since_change = global_episode_idx - last_change
if episodes_since_change < HYSTERESIS_EPISODES:
return current_difficulty, False # too soon to change
# --- apply threshold rules ---
new_difficulty = current_difficulty
if accuracy > HIGH_THRESHOLD:
new_difficulty = min(current_difficulty + 1, MAX_DIFFICULTY)
elif accuracy < LOW_THRESHOLD:
new_difficulty = max(current_difficulty - 1, MIN_DIFFICULTY)
changed = new_difficulty != current_difficulty
if changed:
state.domain_difficulties[domain] = new_difficulty
return new_difficulty, changed
|