Spaces:
Paused
Paused
File size: 8,915 Bytes
31226fd | 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 | """Bounded action-conditioned multistep transition baseline.
The model learns exact discrete transition contexts ``state + action -> next_state``
and can roll them forward over an explicit action sequence. It is deliberately a
statistical reference model, not a neural world model and not accepted evidence
for planning or counterfactual reasoning.
"""
from __future__ import annotations
import copy
import json
from collections import Counter
from dataclasses import dataclass
from typing import Any, Sequence, cast
class MultistepWorldModelError(ValueError):
"""Raised for malformed transition data or rollout requests."""
def _canonical(value: object) -> str:
try:
return json.dumps(
value,
sort_keys=True,
separators=(",", ":"),
ensure_ascii=True,
allow_nan=False,
)
except (TypeError, ValueError) as error:
raise MultistepWorldModelError(
"state and action values must be finite JSON"
) from error
def _state(value: object, name: str) -> dict[str, Any]:
if not isinstance(value, dict):
raise MultistepWorldModelError(f"{name} must be a string-keyed object")
raw = cast(dict[object, object], value)
if not all(isinstance(key, str) for key in raw):
raise MultistepWorldModelError(f"{name} must be a string-keyed object")
typed = cast(dict[str, Any], value)
_canonical(typed)
return copy.deepcopy(typed)
@dataclass(frozen=True, slots=True)
class StateAction:
"""Explicit action label and JSON payload for one transition step."""
action: str
payload: dict[str, Any]
def __post_init__(self) -> None:
if not self.action.strip():
raise MultistepWorldModelError("action must be non-empty")
_state(self.payload, "action payload")
def to_dict(self) -> dict[str, Any]:
return {"action": self.action, "payload": copy.deepcopy(self.payload)}
@dataclass(frozen=True, slots=True)
class RolloutStep:
step: int
action: StateAction
predicted_state: dict[str, Any] | None
support: int
uncertainty: float
@dataclass(frozen=True, slots=True)
class RolloutResult:
initial_state: dict[str, Any]
steps: tuple[RolloutStep, ...]
terminated_early: bool
@property
def final_state(self) -> dict[str, Any] | None:
if not self.steps:
return copy.deepcopy(self.initial_state)
return copy.deepcopy(self.steps[-1].predicted_state)
class ActionConditionedWorldModel:
"""Bounded exact-context transition counts with deterministic rollouts."""
model_version = 1
def __init__(self, *, max_contexts: int = 512) -> None:
if type(max_contexts) is not int or max_contexts <= 0:
raise MultistepWorldModelError("max_contexts must be a positive integer")
self.max_contexts = max_contexts
self._contexts: dict[str, Counter[str]] = {}
@staticmethod
def _context(state: dict[str, Any], action: StateAction) -> str:
return _canonical({"state": state, "action": action.to_dict()})
def update(
self,
state: dict[str, Any],
action: StateAction,
next_state: dict[str, Any],
) -> None:
current = _state(state, "state")
following = _state(next_state, "next_state")
key = self._context(current, action)
label = _canonical(following)
if key not in self._contexts and len(self._contexts) >= self.max_contexts:
del self._contexts[next(iter(self._contexts))]
self._contexts.setdefault(key, Counter[str]())[label] += 1
def predict(
self, state: dict[str, Any], action: StateAction
) -> tuple[dict[str, Any] | None, int, float]:
current = _state(state, "state")
counts = self._contexts.get(self._context(current, action))
if not counts:
return None, 0, 1.0
label = min(counts, key=lambda item: (-counts[item], item))
support = counts[label]
total = sum(counts.values())
return json.loads(label), support, 1.0 - support / total
def rollout(
self,
initial_state: dict[str, Any],
actions: Sequence[StateAction],
*,
max_steps: int = 32,
) -> RolloutResult:
if type(max_steps) is not int or max_steps <= 0:
raise MultistepWorldModelError("max_steps must be a positive integer")
if len(actions) > max_steps:
raise MultistepWorldModelError("action sequence exceeds rollout budget")
initial = _state(initial_state, "initial_state")
current = initial
steps: list[RolloutStep] = []
terminated = False
for index, action in enumerate(actions, start=1):
predicted, support, uncertainty = self.predict(current, action)
steps.append(RolloutStep(index, action, predicted, support, uncertainty))
if predicted is None:
terminated = True
break
current = predicted
return RolloutResult(initial, tuple(steps), terminated)
def counterfactual(
self,
state: dict[str, Any],
actions: Sequence[StateAction],
) -> dict[str, dict[str, Any] | None]:
"""Compare one-step alternatives without claiming causal intervention."""
result: dict[str, dict[str, Any] | None] = {}
for action in actions:
key = _canonical(action.to_dict())
predicted, _, _ = self.predict(state, action)
result[key] = predicted
return result
def state_dict(self) -> dict[str, Any]:
return {
"model_version": self.model_version,
"max_contexts": self.max_contexts,
"context_order": list(self._contexts),
"contexts": {
key: dict(sorted(values.items()))
for key, values in self._contexts.items()
},
"scientific_status": "statistical_multistep_reference_not_neural_evidence",
}
@classmethod
def from_state_dict(cls, state: dict[str, Any]) -> "ActionConditionedWorldModel":
if state.get("model_version") != cls.model_version:
raise MultistepWorldModelError("unsupported multistep world-model version")
try:
model = cls(max_contexts=state["max_contexts"])
raw_contexts = state["contexts"]
raw_order = state["context_order"]
if not isinstance(raw_contexts, dict) or not isinstance(raw_order, list):
raise MultistepWorldModelError("invalid contexts or context order")
contexts = cast(dict[object, object], raw_contexts)
order = cast(list[object], raw_order)
if len(order) != len(contexts) or set(order) != set(contexts):
raise MultistepWorldModelError("context order does not match contexts")
if len(order) > model.max_contexts:
raise MultistepWorldModelError("stored contexts exceed capacity")
for raw_key in order:
if not isinstance(raw_key, str):
raise MultistepWorldModelError("invalid transition context")
raw_counts = contexts.get(raw_key)
if not isinstance(raw_counts, dict):
raise MultistepWorldModelError("invalid transition context")
count_items = cast(dict[object, object], raw_counts)
counts: Counter[str] = Counter()
for raw_label, raw_count in count_items.items():
if (
not isinstance(raw_label, str)
or type(raw_count) is not int
or raw_count <= 0
):
raise MultistepWorldModelError("invalid transition count")
label = raw_label
count = raw_count
decoded = json.loads(label)
if _canonical(_state(decoded, "stored next_state")) != label:
raise MultistepWorldModelError(
"non-canonical stored next_state"
)
counts[label] = count
if not counts:
raise MultistepWorldModelError(
"transition context may not be empty"
)
model._contexts[raw_key] = counts
return model
except (KeyError, TypeError, json.JSONDecodeError) as error:
if isinstance(error, MultistepWorldModelError):
raise
raise MultistepWorldModelError(
"malformed multistep world-model state"
) from error
__all__ = [
"ActionConditionedWorldModel",
"MultistepWorldModelError",
"RolloutResult",
"RolloutStep",
"StateAction",
]
|