File size: 15,146 Bytes
455fdee 8bdd2de 455fdee | 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 | """
Experience Replay — Trajectory storage and retrieval.
Stores completed trajectories with their scores and supports retrieval
ranked by a combination of:
1. Semantic similarity (embedding-based) — from MemRL (arxiv:2601.03192)
2. Learned Q-value utility scores — from REMEMBERER (arxiv:2306.07929)
The two-phase retrieval (recall by similarity → re-rank by Q-value) separates
"semantically similar" from "functionally useful" — a key insight from MemRL.
This module is the "database" — it stores but doesn't analyze. The Optimizer
module reads from here and writes heuristics back.
"""
from __future__ import annotations
import json
import logging
import math
import os
import time
from pathlib import Path
from typing import Any
from purpose_agent.types import (
Heuristic,
MemoryRecord,
MemoryTier,
Trajectory,
TrajectoryStep,
)
logger = logging.getLogger(__name__)
class ExperienceReplay:
"""
Experience Replay buffer with two-phase retrieval.
Phase 1 (Recall): Retrieve top-k records by semantic similarity to query
Phase 2 (Re-rank): Re-order by learned Q-value utility scores
The buffer supports:
- Adding trajectories with automatic scoring
- Retrieving similar past experiences for the Actor's context
- Q-value updates after heuristics are applied (Bellman-style)
- Persistence to disk (JSON)
- Capacity management (evict lowest Q-value records when full)
Args:
capacity: Maximum number of records to store
similarity_weight: λ in retrieval score = λ·similarity + (1-λ)·q_value
persistence_path: If set, auto-save/load buffer to this file
"""
def __init__(
self,
capacity: int = 500,
similarity_weight: float = 0.6,
persistence_path: str | Path | None = None,
):
self.capacity = capacity
self.similarity_weight = similarity_weight
self.persistence_path = Path(persistence_path) if persistence_path else None
self.records: list[MemoryRecord] = []
# Load from disk if available
if self.persistence_path and self.persistence_path.exists():
self._load()
# ------------------------------------------------------------------
# Core Operations
# ------------------------------------------------------------------
def add(self, trajectory: Trajectory) -> MemoryRecord:
"""
Add a completed trajectory to the buffer.
Automatically computes a task embedding (simple TF-IDF-style hash)
and initial Q-value based on trajectory performance.
"""
# Compute initial Q-value from trajectory performance
initial_q = self._compute_initial_q(trajectory)
record = MemoryRecord(
trajectory=trajectory,
heuristics=[],
task_embedding=self._compute_embedding(trajectory.task_description),
retrieval_q_value=initial_q,
)
# Capacity management: evict lowest Q-value if full
if len(self.records) >= self.capacity:
self._evict()
self.records.append(record)
logger.info(
f"Experience Replay: Added trajectory '{trajectory.id}' "
f"(q={initial_q:.3f}, steps={len(trajectory.steps)}, "
f"Σreward={trajectory.cumulative_reward:.2f})"
)
if self.persistence_path:
self._save()
return record
def retrieve(
self,
query: str,
top_k: int = 5,
min_q_value: float = 0.0,
) -> list[MemoryRecord]:
"""
Two-phase retrieval (per MemRL arxiv:2601.03192):
Phase 1: Recall candidates by semantic similarity
Phase 2: Re-rank by Q-value utility
Returns top-k records sorted by combined score.
"""
if not self.records:
return []
query_embedding = self._compute_embedding(query)
# Phase 1: Compute similarity scores for all records
scored: list[tuple[float, MemoryRecord]] = []
for record in self.records:
if record.retrieval_q_value < min_q_value:
continue
sim = self._cosine_similarity(
query_embedding, record.task_embedding or []
)
# Phase 2: Combined score
combined = (
self.similarity_weight * sim
+ (1 - self.similarity_weight) * record.retrieval_q_value
)
scored.append((combined, record))
# Sort descending by combined score
scored.sort(key=lambda x: -x[0])
results = [record for _, record in scored[:top_k]]
logger.debug(
f"Experience Replay: Retrieved {len(results)} records for query "
f"(top score={scored[0][0]:.3f})" if scored else "no records"
)
return results
def update_q_value(
self,
record_id: str,
reward: float,
alpha: float = 0.1,
) -> None:
"""
Update a record's retrieval Q-value using Monte Carlo update.
Q_new = Q_old + α * (reward - Q_old)
From REMEMBERER (arxiv:2306.07929): α = 1/N where N = number of
updates. We use a fixed α for simplicity; override for REMEMBERER-exact.
"""
for record in self.records:
if record.id == record_id:
old_q = record.retrieval_q_value
record.retrieval_q_value += alpha * (reward - old_q)
record.retrieval_q_value = max(0.0, min(1.0, record.retrieval_q_value))
logger.debug(
f"Experience Replay: Q-value update for {record_id}: "
f"{old_q:.3f} → {record.retrieval_q_value:.3f}"
)
if self.persistence_path:
self._save()
return
logger.warning(f"Experience Replay: Record {record_id} not found for Q-update")
def attach_heuristics(
self, record_id: str, heuristics: list[Heuristic]
) -> None:
"""Attach extracted heuristics to a memory record."""
for record in self.records:
if record.id == record_id:
record.heuristics = heuristics
if self.persistence_path:
self._save()
return
# ------------------------------------------------------------------
# Statistics & Queries
# ------------------------------------------------------------------
def get_top_trajectories(
self,
n: int = 10,
min_success_rate: float = 0.5,
) -> list[Trajectory]:
"""Get the n best trajectories by cumulative reward."""
candidates = [
r.trajectory for r in self.records
if r.trajectory.success_rate >= min_success_rate
]
candidates.sort(key=lambda t: -t.cumulative_reward)
return candidates[:n]
def get_all_heuristics(self, tier: MemoryTier | None = None) -> list[Heuristic]:
"""Get all extracted heuristics, optionally filtered by tier."""
heuristics = []
for record in self.records:
for h in record.heuristics:
if tier is None or h.tier == tier:
heuristics.append(h)
return heuristics
@property
def size(self) -> int:
return len(self.records)
def clear(self) -> None:
"""Reset the replay buffer. Removes all records and persists the empty state."""
self.records.clear()
if self.persistence_path:
self._save()
logger.info("Experience Replay: cleared all records")
@property
def stats(self) -> dict[str, Any]:
if not self.records:
return {"size": 0}
q_values = [r.retrieval_q_value for r in self.records]
rewards = [r.trajectory.cumulative_reward for r in self.records]
return {
"size": len(self.records),
"avg_q_value": sum(q_values) / len(q_values),
"max_q_value": max(q_values),
"avg_cumulative_reward": sum(rewards) / len(rewards),
"total_heuristics": sum(len(r.heuristics) for r in self.records),
}
# ------------------------------------------------------------------
# Embedding & Similarity (lightweight — no external deps)
# ------------------------------------------------------------------
@staticmethod
def _compute_embedding(text: str) -> list[float]:
"""
Lightweight text embedding using character n-gram hashing.
This is intentionally simple — for production, swap in a real
embedding model (sentence-transformers, OpenAI embeddings, etc.).
To use real embeddings, subclass ExperienceReplay and override
_compute_embedding() and _cosine_similarity().
"""
# Character trigram hashing into a fixed-size vector
dim = 128
vec = [0.0] * dim
text_lower = text.lower()
for i in range(len(text_lower) - 2):
trigram = text_lower[i:i + 3]
h = hash(trigram) % dim
vec[h] += 1.0
# L2 normalize
magnitude = math.sqrt(sum(x * x for x in vec))
if magnitude > 0:
vec = [x / magnitude for x in vec]
return vec
@staticmethod
def _cosine_similarity(a: list[float], b: list[float]) -> float:
"""Cosine similarity between two vectors."""
if not a or not b or len(a) != len(b):
return 0.0
dot = sum(x * y for x, y in zip(a, b))
mag_a = math.sqrt(sum(x * x for x in a))
mag_b = math.sqrt(sum(x * x for x in b))
if mag_a == 0 or mag_b == 0:
return 0.0
return dot / (mag_a * mag_b)
# ------------------------------------------------------------------
# Initial Q-Value Estimation
# ------------------------------------------------------------------
@staticmethod
def _compute_initial_q(trajectory: Trajectory) -> float:
"""
Compute initial Q-value from trajectory performance.
Uses a combination of:
- Success rate (fraction of steps that improved state)
- Total delta (net improvement)
- Trajectory length efficiency (shorter = better for same delta)
"""
if not trajectory.steps:
return 0.3 # Uninformative prior
success_rate = trajectory.success_rate
total_delta = trajectory.total_delta
length = len(trajectory.steps)
# Normalize total_delta to 0-1 (assuming max meaningful delta is ~10)
delta_normalized = max(0.0, min(1.0, total_delta / 10.0))
# Efficiency bonus: more progress per step = higher Q
efficiency = delta_normalized / max(length, 1)
q = 0.4 * success_rate + 0.4 * delta_normalized + 0.2 * min(efficiency * 5, 1.0)
return max(0.0, min(1.0, q))
# ------------------------------------------------------------------
# Capacity Management
# ------------------------------------------------------------------
def _evict(self) -> None:
"""Evict the lowest Q-value record."""
if not self.records:
return
worst = min(self.records, key=lambda r: r.retrieval_q_value)
self.records.remove(worst)
logger.debug(
f"Experience Replay: Evicted record {worst.id} "
f"(q={worst.retrieval_q_value:.3f})"
)
# ------------------------------------------------------------------
# Persistence
# ------------------------------------------------------------------
def _save(self) -> None:
"""Save buffer to disk as JSON."""
if not self.persistence_path:
return
self.persistence_path.parent.mkdir(parents=True, exist_ok=True)
data = []
for record in self.records:
data.append({
"id": record.id,
"retrieval_q_value": record.retrieval_q_value,
"task_embedding": record.task_embedding,
"trajectory": {
"id": record.trajectory.id,
"task_description": record.trajectory.task_description,
"purpose": record.trajectory.purpose,
"created_at": record.trajectory.created_at,
"cumulative_reward": record.trajectory.cumulative_reward,
"total_delta": record.trajectory.total_delta,
"success_rate": record.trajectory.success_rate,
"num_steps": len(record.trajectory.steps),
},
"heuristics": [
{
"id": h.id,
"pattern": h.pattern,
"strategy": h.strategy,
"steps": h.steps,
"tier": h.tier.value,
"q_value": h.q_value,
"times_used": h.times_used,
"times_succeeded": h.times_succeeded,
}
for h in record.heuristics
],
})
with open(self.persistence_path, "w") as f:
json.dump(data, f, indent=2, default=str)
def _load(self) -> None:
"""Load buffer from disk."""
if not self.persistence_path or not self.persistence_path.exists():
return
try:
with open(self.persistence_path) as f:
data = json.load(f)
for entry in data:
traj_data = entry["trajectory"]
trajectory = Trajectory(
task_description=traj_data["task_description"],
purpose=traj_data["purpose"],
id=traj_data["id"],
created_at=traj_data.get("created_at", time.time()),
)
heuristics = [
Heuristic(
id=h["id"],
pattern=h["pattern"],
strategy=h["strategy"],
steps=h["steps"],
tier=MemoryTier(h["tier"]),
q_value=h["q_value"],
times_used=h.get("times_used", 0),
times_succeeded=h.get("times_succeeded", 0),
)
for h in entry.get("heuristics", [])
]
record = MemoryRecord(
id=entry["id"],
trajectory=trajectory,
heuristics=heuristics,
task_embedding=entry.get("task_embedding"),
retrieval_q_value=entry.get("retrieval_q_value", 0.5),
)
self.records.append(record)
logger.info(f"Experience Replay: Loaded {len(self.records)} records from disk")
except Exception as e:
logger.error(f"Experience Replay: Failed to load from disk: {e}")
|