Spaces:
Sleeping
Sleeping
File size: 9,462 Bytes
0fff343 | 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 | """Typed GP loop for engine_v2.
Init via ramped half-and-half. Tournament selection. Elitism. Subtree
crossover with strict type matching. Subtree + point mutation. A single
seed governs all RNG (init, selection, crossover, mutation, FeatureSet
sampling). The fitness cache is keyed by ``program.repr_typed()`` so
elites that carry across generations don't recompute.
"""
from __future__ import annotations
import random
from typing import Callable, Sequence
import numpy as np
import pandas as pd
from engine_v2.fitness import V2Objective, fitness_fn
from engine_v2.nodes import ExecContext, Node
from engine_v2.synthesize import (
crossover,
mutate,
ramped_population,
)
def _assign_id(program: Node, generation: int, idx: int) -> str:
return f"g{generation}c{idx}"
def _tournament_select(
rng: random.Random,
population: list[Node],
fitnesses: list[float],
k: int,
) -> int:
contenders = rng.sample(range(len(population)), k)
return max(contenders, key=lambda i: fitnesses[i])
def run_gp_v2(
ctx_train: ExecContext,
y_train: np.ndarray | None,
pool: Sequence[str],
*,
objective: V2Objective,
population_size: int = 150,
n_generations: int = 30,
tournament_k: int = 3,
elitism: int = 5,
p_mutate: float = 0.7,
lambda_size: float = 0.005,
cv_folds: int = 5,
seed: int = 42,
max_depth: int = 4,
max_genes_per_set: int = 8,
max_nodes: int = 64,
coherence_weight: float = 0.0,
immigrant_fraction: float = 0.0,
rates_override: dict | None = None,
scalar_share_override: float | None = None,
on_generation: Callable[[dict], None] | None = None,
) -> tuple[list[dict], Node, float]:
"""Returns ``(generation_log, winner, winner_fitness)``.
Each ``generation_log`` entry is a complete snapshot of the
population: every candidate with its id, program_repr, fitness,
survived flag, n_nodes, and depth. Top-12 are also exposed inline
via ``top_candidates`` for the SSE stream's bandwidth budget.
"""
py_rng = random.Random(seed)
# Per-objective synthesis overrides (e.g. UNSUP forbids Scalar roots
# and Associate / Effect / FitApply — Vector-only programs). The
# caller may also pass `rates_override` (e.g. to gate Search off
# per-run); those keys win over the objective's overrides which
# win over DEFAULT_RATES inside synthesize. ``scalar_share`` lives
# on its own param (it's a single float, not a per-operator rate);
# an explicit `scalar_share_override` wins over the objective's
# value.
overrides = objective.synthesis_overrides() if hasattr(objective, "synthesis_overrides") else {}
objective_rates = overrides.get("rates") or {}
objective_scalar_share = overrides.get("scalar_share")
effective_scalar_share = (
scalar_share_override
if scalar_share_override is not None
else objective_scalar_share
)
merged_rates_override: dict | None
if rates_override is None and not objective_rates:
merged_rates_override = None
else:
merged_rates_override = {
**(objective_rates or {}),
**(rates_override or {}),
}
population: list[Node] = ramped_population(
py_rng, pool,
n=population_size,
objective_target=objective.target,
max_depth=max_depth,
max_genes_per_set=max_genes_per_set,
**(
{"rates": merged_rates_override} if merged_rates_override is not None else {}
),
**(
{"scalar_share": effective_scalar_share}
if effective_scalar_share is not None else {}
),
)
ids: list[str] = [_assign_id(p, 0, i) for i, p in enumerate(population)]
parents: list[list[str]] = [[] for _ in population]
fitness_cache: dict[str, float] = {}
def evaluate(pop: list[Node]) -> list[float]:
out = []
for prog in pop:
sig = prog.repr_typed()
if sig not in fitness_cache:
fitness_cache[sig] = fitness_fn(
prog, ctx_train, y_train,
objective=objective,
lambda_size=lambda_size,
n_folds=cv_folds,
random_state=seed,
coherence_weight=coherence_weight,
)
out.append(fitness_cache[sig])
return out
worst = objective.worst_score()
def _finite(x: float) -> float:
return float(x) if np.isfinite(x) else worst
log: list[dict] = []
best_overall: Node | None = None
best_overall_fit = -np.inf
for gen in range(n_generations):
fitnesses = evaluate(population)
ranked = sorted(
range(len(population)),
key=lambda i: fitnesses[i],
reverse=True,
)
elite_ix = set(ranked[:elitism])
candidates_payload = [
{
"id": ids[i],
"program_repr": population[i].repr_typed(),
"fitness": _finite(fitnesses[i]),
"n_nodes": int(population[i].node_count()),
"depth": int(population[i].depth()),
"gene_ids": list(population[i].feature_ids()),
"parents": list(parents[i]),
"survived": i in elite_ix,
}
for i in ranked
]
top_payload = candidates_payload[:12]
# Median computed over the finite values only, so a generation
# where many programs are degenerate doesn't collapse to -inf.
finite_vals = [f for f in fitnesses if np.isfinite(f)]
median_fit = float(np.median(finite_vals)) if finite_vals else worst
entry = {
"generation": gen,
"best_fitness": _finite(fitnesses[ranked[0]]),
"median_fitness": median_fit,
"elitism": elitism,
"population_size": len(population),
"top_candidates": top_payload,
"candidates": candidates_payload,
}
log.append(entry)
if on_generation is not None:
on_generation(entry)
if fitnesses[ranked[0]] > best_overall_fit:
best_overall_fit = fitnesses[ranked[0]]
best_overall = population[ranked[0]]
if gen == n_generations - 1:
break
# ----- Build next generation -----
new_pop: list[Node] = []
new_ids: list[str] = []
new_parents: list[list[str]] = []
# Elites carry over unchanged.
for ei, idx in enumerate(ranked[:elitism]):
new_pop.append(population[idx]) # share the node — it's evaluated.
new_ids.append(_assign_id(population[idx], gen + 1, ei))
new_parents.append([ids[idx]])
i_offset = elitism
# Random immigrants — fresh programs drawn from ramped_population
# using the same grammar / objective / depth constraints as init.
# They displace offspring slots (never elites), counter premature
# convergence, and are gated to immigrant_fraction > 0 so existing
# runs are byte-for-byte unchanged. Rounding: round(frac * pop),
# capped so it never pushes past the budget after elites.
n_immigrants = (
int(round(immigrant_fraction * population_size))
if immigrant_fraction > 0.0 else 0
)
n_immigrants = max(0, min(n_immigrants, population_size - len(new_pop)))
if n_immigrants > 0:
fresh = ramped_population(
py_rng, pool,
n=n_immigrants,
objective_target=objective.target,
max_depth=max_depth,
max_genes_per_set=max_genes_per_set,
**(
{"rates": merged_rates_override} if merged_rates_override is not None else {}
),
**(
{"scalar_share": effective_scalar_share}
if effective_scalar_share is not None else {}
),
)
for child in fresh:
new_pop.append(child)
new_ids.append(_assign_id(child, gen + 1, i_offset))
new_parents.append([]) # immigrant — no parents
i_offset += 1
while len(new_pop) < population_size:
i1 = _tournament_select(py_rng, population, fitnesses, tournament_k)
i2 = _tournament_select(py_rng, population, fitnesses, tournament_k)
child = crossover(
py_rng, population[i1], population[i2],
max_depth=max_depth, max_nodes=max_nodes,
)
child = mutate(
py_rng, child, pool,
objective_target=objective.target,
p_mut=p_mutate,
max_depth=max_depth,
max_genes_per_set=max_genes_per_set,
max_nodes=max_nodes,
**(
{"rates": merged_rates_override}
if merged_rates_override is not None else {}
),
)
new_pop.append(child)
new_ids.append(_assign_id(child, gen + 1, i_offset))
new_parents.append([ids[i1], ids[i2]])
i_offset += 1
population, ids, parents = new_pop, new_ids, new_parents
assert best_overall is not None
return log, best_overall, float(best_overall_fit)
|