Spaces:
Sleeping
Sleeping
File size: 20,454 Bytes
b0b140b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 | """LandscapeForge OpenEnv environment — OptCoder REPL (Phase C).
For v1 we ship OptCoder-only: LandscapeForge is a fixed template picker
controlled by the env itself (uniform random over the tier menu). The agent
acting through OpenEnv is OptCoder.
Each `reset()` samples a new landscape from the current tier. Each `step()`
executes one OptCoder action (run_baseline / draft / inspect / commit),
mutates env state, and returns an Observation reflecting the new state.
Episode ends when OptCoder commits or budget is exhausted.
"""
from __future__ import annotations
from typing import Any, Optional
from uuid import uuid4
import numpy as np
from openenv.core.env_server.interfaces import Environment
from openenv.core.env_server.types import State
try:
from ..models import (
ACTION_COSTS,
LandscapeforgeAction,
LandscapeforgeObservation,
)
from ..landscapes import (
TIER_MENU,
Landscape,
build_landscape,
structural_hints,
)
from ..reference_optimizers import run_baseline as run_reference_baseline
from ..reference_optimizers import tune_adam_lr
from ..sandbox import SandboxError, compile_optimizer
from ..arena import ArenaResult, auto_test_draft, run_arena
from ..rewards import ast_novelty_score, compute_optcoder_reward, compute_step_reward
except ImportError:
# Running from repo root or package layout quirks
from models import ( # type: ignore
ACTION_COSTS,
LandscapeforgeAction,
LandscapeforgeObservation,
)
from landscapes import ( # type: ignore
TIER_MENU,
Landscape,
build_landscape,
structural_hints,
)
from reference_optimizers import run_baseline as run_reference_baseline # type: ignore
from reference_optimizers import tune_adam_lr # type: ignore
from sandbox import SandboxError, compile_optimizer # type: ignore
from arena import ArenaResult, auto_test_draft, run_arena # type: ignore
from rewards import ast_novelty_score, compute_optcoder_reward, compute_step_reward # type: ignore
BUDGET_TOTAL = 12
ARENA_SEEDS = [101, 202, 303, 404, 505, 606, 707, 808, 909, 1010]
ARENA_STEPS = 200
BASELINE_STEPS = 30 # env-controlled; agent does not choose
# Reference source blobs for AST novelty comparison (short pseudo-implementations).
# Kept minimal — enough to detect "this commit is basically Adam".
_REF_SGD = """
class Optimizer:
def __init__(self, dim): self.lr = 0.01
def step(self, x, f, g): return x - self.lr * g
""".strip()
def _adam_source(lr: float) -> str:
"""Adam reference implementation parameterized by LR.
Used by `_ensure_adam_arena` after LR tuning — the baseline is
Adam-at-best-LR-for-this-landscape, not Adam-at-fixed-default.
"""
return f"""
class Optimizer:
def __init__(self, dim):
self.lr = {lr}
self.b1 = 0.9
self.b2 = 0.999
self.eps = 1e-8
self.m = np.zeros(dim)
self.v = np.zeros(dim)
self.t = 0
def step(self, x, f_val, g):
self.t += 1
self.m = self.b1*self.m + (1-self.b1)*g
self.v = self.b2*self.v + (1-self.b2)*g*g
mh = self.m/(1-self.b1**self.t)
vh = self.v/(1-self.b2**self.t)
return x - self.lr * mh / (np.sqrt(vh) + self.eps)
""".strip()
# Frozen default-LR source used only for AST-novelty comparison (so r_novelty
# measures "structurally different from Adam" regardless of the tuned LR).
_REF_ADAM = _adam_source(0.001)
_REF_MOMENTUM = """
class Optimizer:
def __init__(self, dim):
import numpy as np
self.lr=0.01; self.beta=0.9; self.v = np.zeros(dim)
def step(self, x, f, g):
self.v = self.beta*self.v - self.lr*g
return x + self.v
""".strip()
REFERENCE_SOURCES = [_REF_SGD, _REF_ADAM, _REF_MOMENTUM]
class LandscapeforgeEnvironment(Environment):
"""OptCoder-facing OpenEnv environment.
LandscapeForge is internal (template picker) in v1.
"""
SUPPORTS_CONCURRENT_SESSIONS: bool = True
def __init__(self, tier: str = "T0", seed: int = 0):
self._initial_tier = tier
self._master_rng = np.random.default_rng(seed)
self._reset_count = 0
self._tier = tier
self._state = State(episode_id=str(uuid4()), step_count=0)
# Populated by reset()
self._landscape: Optional[Landscape] = None
self._hints: dict = {}
self._baseline_history: list[dict] = []
self._draft_history: list[dict] = []
self._draft_details: list[list[dict]] = [] # per-draft per-step detail
self._inspect_requests: list[dict] = []
self._current_draft: Optional[str] = None
self._budget_spent: int = 0
self._committed: bool = False
self._final_obs: Optional[LandscapeforgeObservation] = None
# Cache Adam's full arena result per episode (computed lazily, for
# reward normalization via progress-based r_regret). The baseline is
# Adam-at-tuned-LR — per-landscape LR is selected via a short sweep.
self._adam_arena_cache: Optional[ArenaResult] = None
self._adam_tuned_lr: Optional[float] = None
# Stepwise feedback log (PBS delta + compile penalty). This is shown to
# the LLM in the observation so it can course-correct mid-episode, but
# NEVER added to the training scalar — final reward is purely terminal
# arena reward (§9.1) for robustness against reward hacking.
self._step_feedback_log: list[dict] = []
# ---------- OpenEnv API ----------
def reset(self) -> LandscapeforgeObservation:
self._reset_count += 1
self._state = State(episode_id=str(uuid4()), step_count=0)
# Pick a landscape from the current tier's menu.
menu = TIER_MENU[self._tier]
template = str(self._master_rng.choice(menu))
dim = int(self._master_rng.integers(2, 6)) # small dims for v1
params = self._sample_params(template)
self._landscape = build_landscape(
template=template, dim=dim, params=params,
rng=np.random.default_rng(int(self._master_rng.integers(0, 2**31))),
)
self._hints = structural_hints(
self._landscape,
rng=np.random.default_rng(int(self._master_rng.integers(0, 2**31))),
)
# Wipe REPL state
self._baseline_history = []
self._draft_history = []
self._draft_details = []
self._inspect_requests = []
self._current_draft = None
self._budget_spent = 0
self._committed = False
self._final_obs = None
self._adam_arena_cache = None
self._adam_tuned_lr = None
self._step_feedback_log = []
return self._make_observation(
last_kind=None, last_result={"reset": True}, done=False, reward=0.0,
)
def step(self, action: LandscapeforgeAction) -> LandscapeforgeObservation: # type: ignore[override]
if self._landscape is None:
raise RuntimeError("step() called before reset()")
if self._committed:
# Episode already done; return terminal obs.
assert self._final_obs is not None
return self._final_obs
self._state.step_count += 1
cost = ACTION_COSTS[action.kind]
# Charge budget first so over-limit actions are rejected.
if self._budget_spent + cost > BUDGET_TOTAL and action.kind != "commit":
return self._force_commit(reason="budget_exhausted")
self._budget_spent += cost
# Snapshot draft history for PBS computation
prev_draft_history_snapshot = list(self._draft_history)
if action.kind == "run_baseline":
result = self._do_run_baseline(action)
elif action.kind == "draft":
result = self._do_draft(action)
elif action.kind == "inspect":
result = self._do_inspect(action)
elif action.kind == "commit":
return self._do_commit()
else:
raise ValueError(f"Unknown action kind: {action.kind}")
# Compute stepwise FEEDBACK (NOT reward). Signals the LLM can use to
# course-correct mid-episode — exposed through last_action_result.
# Explicitly NOT summed into training reward; terminal arena reward
# is the only signal GRPO sees (robust against reward hacking).
step_feedback = compute_step_reward(
prev_draft_history=prev_draft_history_snapshot,
new_draft_history=self._draft_history,
action_kind=action.kind,
action_result=result,
)
if step_feedback["breakdown"]:
entry = {
"turn": self._state.step_count,
"action_kind": action.kind,
**step_feedback["breakdown"],
}
self._step_feedback_log.append(entry)
# Surface on this turn's action result so the LLM sees it immediately.
result = {**result, "feedback": step_feedback["breakdown"]}
# Check if budget now exhausted; if so, auto-commit.
if self._budget_spent >= BUDGET_TOTAL:
return self._force_commit(reason="budget_exhausted")
return self._make_observation(
last_kind=action.kind, last_result=result,
done=False, reward=0.0, # no reward on non-terminal steps
)
@property
def state(self) -> State:
return self._state
# ---------- Action handlers ----------
def _do_run_baseline(self, action: LandscapeforgeAction) -> dict:
assert self._landscape is not None
# Fixed init AND fixed step count for baseline comparability across
# episodes and rollouts (important for GRPO group-relative advantages).
rng = np.random.default_rng(42)
x0 = rng.normal(0.0, 0.5, size=self._landscape.dim)
result = run_reference_baseline(
name=action.baseline_name, f=self._landscape.f, grad=self._landscape.grad,
x0=x0, steps=BASELINE_STEPS,
)
self._baseline_history.append(result)
return {
"baseline_index": len(self._baseline_history) - 1,
"name": result["name"],
"n_steps": len(result["trajectory"]),
"final_f": (result["trajectory"][-1]["f"]
if result["trajectory"] and result["trajectory"][-1]["f"] is not None
else None),
}
def _do_draft(self, action: LandscapeforgeAction) -> dict:
assert self._landscape is not None
code = action.code or ""
self._current_draft = code
try:
opt = compile_optimizer(code, dim=self._landscape.dim)
except SandboxError as e:
# Record failed draft; still counts toward history for inspect.
self._draft_history.append({
"code": code,
"compile_error": str(e),
"summary": {"converged": False, "diverged": True, "error": str(e),
"final_f": None, "step_of_min": None, "min_f": None},
})
self._draft_details.append([])
return {"draft_index": len(self._draft_history) - 1,
"compile_error": str(e), "summary": None}
test = auto_test_draft(opt, self._landscape, seed=0, steps=20)
self._draft_history.append({
"code": code,
"compile_error": None,
"summary": test["summary"],
})
self._draft_details.append(test["detail"])
return {"draft_index": len(self._draft_history) - 1,
"compile_error": None, "summary": test["summary"]}
def _do_inspect(self, action: LandscapeforgeAction) -> dict:
idx = action.draft_idx
if idx is None or idx < 0 or idx >= len(self._draft_details):
return {"error": f"draft_idx {idx} out of range (have {len(self._draft_details)} drafts)"}
detail = self._draft_details[idx]
start = action.step_range_start
end = min(action.step_range_end, len(detail))
sliced = detail[start:end]
record = {
"draft_idx": idx,
"step_range": [start, end],
"detail": sliced,
}
self._inspect_requests.append(record)
return {"draft_idx": idx, "step_range": [start, end], "n_steps": len(sliced)}
def _do_commit(self) -> LandscapeforgeObservation:
return self._finalize_episode(reason="commit")
def _force_commit(self, reason: str) -> LandscapeforgeObservation:
return self._finalize_episode(reason=reason)
# ---------- Episode finalization ----------
def _finalize_episode(self, reason: str) -> LandscapeforgeObservation:
assert self._landscape is not None
self._committed = True
# Need a current_draft. If none, produce a worst-case result.
if not self._current_draft:
result = {
"reason": reason,
"no_draft": True,
"final_regret": 1.0,
}
r_total = -1.0
breakdown = {"no_draft": 1.0}
obs = self._make_observation(
last_kind="commit", last_result=result,
done=True, reward=r_total,
)
obs.committed = True
obs.final_regret = 1.0
obs.r_optcoder = r_total
obs.r_optcoder_breakdown = breakdown
self._final_obs = obs
return obs
# Full Phase-D arena eval
try:
opt = compile_optimizer(self._current_draft, dim=self._landscape.dim)
arena = run_arena(opt, self._landscape, seeds=ARENA_SEEDS, steps=ARENA_STEPS)
except SandboxError as e:
# Committed code fails to compile -> worst-case result
arena = ArenaResult(
initial_values=[1.0] * len(ARENA_SEEDS),
final_values=[float("nan")] * len(ARENA_SEEDS),
crashed=[True] * len(ARENA_SEEDS),
trajectories=[[] for _ in ARENA_SEEDS],
)
# Adam baseline arena for normalization (always run for reward stability).
adam_arena = self._ensure_adam_arena()
novelty = ast_novelty_score(self._current_draft, REFERENCE_SOURCES)
# Convergence step: first seed's trajectory, first step where f < 0.01 * f0
convergence_step = self._compute_convergence_step(arena)
reward = compute_optcoder_reward(
arena=arena,
adam_arena=adam_arena,
actions_used_cost=self._budget_spent,
budget_total=BUDGET_TOTAL,
novelty_score=novelty,
convergence_step=convergence_step,
arena_steps=ARENA_STEPS,
)
result = {
"reason": reason,
"my_mean_progress": arena.mean_progress,
"adam_mean_progress": adam_arena.mean_progress,
"adam_tuned_lr": self._adam_tuned_lr,
"speedup_vs_adam": reward.breakdown.get("speedup_vs_adam"),
"crash_fraction": arena.crash_fraction,
"novelty_score": novelty,
"convergence_step": convergence_step,
}
obs = self._make_observation(
last_kind="commit", last_result=result,
done=True, reward=reward.r_total,
)
obs.committed = True
# `final_regret` is reinterpreted (no f_min dependency): Adam-shortfall
# in [0, 1]. 0 = matched or beat Adam's descent; 1 = made zero progress
# while Adam descended normally. Capped at 1.
speedup = reward.breakdown.get("speedup_vs_adam", 0.0)
obs.final_regret = float(max(0.0, min(1.0, 1.0 - speedup)))
obs.r_optcoder = reward.r_total
obs.r_optcoder_breakdown = reward.breakdown
self._final_obs = obs
return obs
# ---------- Helpers ----------
def _make_observation(self, last_kind: Optional[str], last_result: dict,
done: bool, reward: float) -> LandscapeforgeObservation:
assert self._landscape is not None
return LandscapeforgeObservation(
landscape_description=self._landscape.description,
dim=self._landscape.dim,
structural_hints=self._hints,
baseline_history=self._serialize_baseline_history(),
draft_history=self._serialize_draft_history(),
inspect_requests=list(self._inspect_requests),
current_draft=self._current_draft,
budget_remaining=BUDGET_TOTAL - self._budget_spent,
last_action_kind=last_kind,
last_action_result=last_result,
done=done,
reward=reward,
)
def _serialize_baseline_history(self) -> list[dict]:
# Trim trajectory to summary-friendly size (every step, x as list).
return [
{"name": b["name"], "trajectory": b["trajectory"]}
for b in self._baseline_history
]
def _serialize_draft_history(self) -> list[dict]:
# For the observation we include code + summary per draft.
return [
{"code": d["code"], "summary": d["summary"], "compile_error": d["compile_error"]}
for d in self._draft_history
]
def _sample_params(self, template: str) -> dict:
rng = self._master_rng
if template == "quadratic":
# T0 uses cond up to 100; T1 up to 1000; T2 higher.
cap = {"T0": 100.0, "T1": 1000.0, "T2": 10_000.0}[self._tier]
return {"cond": float(rng.uniform(1.0, cap))}
if template == "gaussian_mix":
return {
"k": int(rng.integers(2, 6)),
"sigma": float(rng.uniform(0.3, 1.0)),
"spread": float(rng.uniform(1.0, 4.0)),
}
if template == "huber":
return {"delta": float(rng.uniform(0.5, 2.0))}
return {}
def _ensure_adam_arena(self) -> ArenaResult:
"""Build the Adam baseline, FAIRLY — LR is tuned per landscape before
running the arena. The tuning uses a short 30-step sweep on a dedicated
seed (not one of the arena seeds) to avoid overfitting.
Cached per episode in `_adam_arena_cache`. Tuned LR is stored in
`_adam_tuned_lr` for logging / demo surfacing.
"""
if self._adam_arena_cache is not None:
return self._adam_arena_cache
assert self._landscape is not None
try:
# Tune LR on seed 0 (not in ARENA_SEEDS), 30-step sweep.
tune_rng = np.random.default_rng(0)
tune_x0 = tune_rng.normal(0.0, 0.5, size=self._landscape.dim)
best_lr = tune_adam_lr(
f=self._landscape.f, grad=self._landscape.grad,
x0=tune_x0, sweep_steps=30,
)
self._adam_tuned_lr = best_lr
adam_opt = compile_optimizer(_adam_source(best_lr), dim=self._landscape.dim)
self._adam_arena_cache = run_arena(
adam_opt, self._landscape,
seeds=ARENA_SEEDS, steps=ARENA_STEPS,
)
except Exception:
self._adam_tuned_lr = None
self._adam_arena_cache = ArenaResult(
initial_values=[1.0] * len(ARENA_SEEDS),
final_values=[1.0] * len(ARENA_SEEDS),
crashed=[True] * len(ARENA_SEEDS),
trajectories=[[] for _ in ARENA_SEEDS],
)
return self._adam_arena_cache
def _compute_convergence_step(self, arena) -> Optional[int]:
"""First step on first seed where f < 1% of initial f."""
if not arena.trajectories or not arena.trajectories[0]:
return None
traj = arena.trajectories[0]
if not traj:
return None
f0 = traj[0]["f"]
if f0 <= 0:
return None
threshold = 0.01 * f0
for t, snap in enumerate(traj):
if snap["f"] < threshold:
return t
return None
# ---------- Tier advancement API (used by trainer, not agent) ----------
def advance_tier(self, new_tier: str) -> None:
if new_tier not in TIER_MENU:
raise ValueError(f"Unknown tier {new_tier}")
self._tier = new_tier
|