Spaces:
Running
Running
File size: 11,214 Bytes
3d46076 | 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 | import os
import json
import hashlib
import numpy as np
from typing import Dict, Any, List, Optional
from src.connectome.types import ConnectomeGraph
from src.brain.runtime import BrainRuntime
from src.evolution.mutation import StructuralMutator
from src.common.determinism import deterministic_id
class EvolutionScheduler:
"""
Parallel candidate evolution scheduler.
Evaluates candidate brain variants on benchmark battery, records comprehensive metadata
(parent ID, generation, mutation set, seed, score, hashes), and supports rollback.
All identity is deterministic (R8): no uuid4 in scientific records.
"""
def __init__(self, base_graph: ConnectomeGraph, history_file: str = "diagnostics/evolution_history.json",
seed: int = 100):
self.current_graph = base_graph
self.history_file = history_file
self.seed = int(seed)
self.generation = 0
self.current_brain_id = f"brain_gen0_{deterministic_id('scheduler', str(self.seed))[:8]}"
self.history: List[Dict[str, Any]] = []
os.makedirs(os.path.dirname(history_file), exist_ok=True)
def _hash_graph(self, graph: ConnectomeGraph) -> str:
h = hashlib.sha256()
h.update(graph.weights.tobytes())
h.update(graph.col_indices.tobytes())
h.update(graph.row_offsets.tobytes())
return h.hexdigest()
def evaluate_candidate(self, graph: ConnectomeGraph, seed: int = 42) -> float:
"""
Standardized fitness benchmark:
Evaluates dynamic responsiveness, sensory sensitivity, and signal-to-noise ratio.
"""
runtime = BrainRuntime(graph, use_gpu=False, enable_plasticity=False, seed=seed)
scores = []
# Run multi-step sensory stimulation and reward trial
for _ in range(3):
sensory = {"visual": np.full(min(64, graph.num_neurons), 1.5, dtype=np.float32)}
runtime.step(sensory_inputs=sensory, reward=0.5)
mean_act = float(np.mean(runtime.state.activations))
spikes = runtime.state.total_spikes
speak_score = float(runtime.state.tool_associations.get("speak", 0.0))
# Fitness combines responsiveness, selectivity, and synaptic connectivity
fitness = (mean_act * 5.0) + (speak_score * 2.0) + (0.005 * min(spikes, 100)) + (0.0001 * graph.num_synapses)
runtime.cleanup()
return float(round(fitness, 4))
def run_generation(
self,
num_candidates: int = 4,
seed: int = 100
) -> Dict[str, Any]:
"""
Executes one evolutionary generation across multiple candidate variants.
Evaluates all variants and selects the best candidate if it improves fitness.
"""
self.generation += 1
mutator = StructuralMutator(seed=seed + self.generation)
baseline_score = self.evaluate_candidate(self.current_graph, seed=seed)
# V4 (mission §22): capture parent identity BEFORE candidate selection so
# an accepted child can never become its own parent in history.
generation_parent_id = self.current_brain_id
print(f"\n--- Generation {self.generation} (Baseline Score: {baseline_score:.4f}) ---")
candidates = []
best_candidate = None
best_score = baseline_score
for c_idx in range(num_candidates):
cand_seed = seed + self.generation * 10 + c_idx
cand_graph = mutator.clone_graph(self.current_graph)
# Apply distinct mutation combinations
mutations = []
if c_idx == 0:
mutations.append(mutator.mutate_weights(cand_graph, rate=0.08, scale=0.15))
elif c_idx == 1:
mutations.append(mutator.prune_synapses(cand_graph, threshold=0.04))
mutations.append(mutator.rewire_synapses(cand_graph, num_new=10))
elif c_idx == 2:
mutations.append(mutator.grow_population(cand_graph, new_neurons_count=4))
mutations.append(mutator.mutate_weights(cand_graph, rate=0.05, scale=0.1))
else:
mutations.append(mutator.rewire_synapses(cand_graph, num_new=15))
mutations.append(mutator.mutate_weights(cand_graph, rate=0.04, scale=0.08))
cand_score = self.evaluate_candidate(cand_graph, seed=cand_seed)
cand_id = (f"brain_gen{self.generation}_c{c_idx}_"
f"{deterministic_id(str(self.seed), str(self.generation), str(c_idx))[:6]}")
cand_hash = self._hash_graph(cand_graph)
accepted = cand_score > best_score
decision_reason = "Improved fitness over baseline" if accepted else "Fitness did not exceed best candidate"
cand_record = {
"candidate_id": cand_id,
"parent_id": self.current_brain_id,
"generation": self.generation,
"seed": cand_seed,
"num_neurons": cand_graph.num_neurons,
"num_synapses": cand_graph.num_synapses,
"mutations": mutations,
"score": cand_score,
"baseline_score": baseline_score,
"accepted": accepted,
"reason": decision_reason,
"sha256": cand_hash
}
candidates.append(cand_record)
print(f"Candidate {c_idx}: Score={cand_score:.4f}, Neurons={cand_graph.num_neurons}, Synapses={cand_graph.num_synapses} "
f"| {'[NEW LEADER]' if accepted else '[REJECTED]'}")
if accepted:
best_score = cand_score
best_candidate = (cand_id, cand_graph, cand_record)
# Selection or Rollback
if best_candidate is not None and best_score > baseline_score:
self.current_brain_id, self.current_graph, best_rec = best_candidate
best_rec["reason"] = "ACCEPTED: Highest fitness in generation"
print(f"-> Selected Candidate {self.current_brain_id} (Score: {best_score:.4f})")
else:
print(f"-> Rollback: Retained parent baseline {self.current_brain_id} (Score: {baseline_score:.4f})")
gen_summary = {
"generation": self.generation,
"parent_id": generation_parent_id,
"child_id": self.current_brain_id,
"baseline_score": baseline_score,
"best_score": best_score,
"improved": best_score > baseline_score,
"candidates": candidates
}
self.history.append(gen_summary)
self.save_history()
return gen_summary
def get_lineage(self) -> Dict[str, Any]:
"""Returns the full evolutionary lineage tree with candidate decisions."""
return {
"current_generation": self.generation,
"current_brain_id": self.current_brain_id,
"current_neurons": self.current_graph.num_neurons,
"current_synapses": self.current_graph.num_synapses,
"history": self.history
}
def save_history(self):
with open(self.history_file, "w", encoding="utf-8") as f:
json.dump(self.get_lineage(), f, indent=2)
def checkpoint_path(self) -> str:
base, _ = os.path.splitext(self.history_file)
return base + ".checkpoint.npz"
def save_checkpoint(self) -> str:
"""Persists full resumable state: current graph, generation, IDs, seed."""
from src.connectome.types import PopulationMetadata, ProvenanceStatus
path = self.checkpoint_path()
os.makedirs(os.path.dirname(os.path.abspath(path)), exist_ok=True)
pop_ser = {}
if self.current_graph.populations is not None:
for name, pm in self.current_graph.populations.populations.items():
pop_ser[name] = {
"indices": [int(x) for x in pm.neuron_indices],
"ids": [int(x) for x in pm.neuron_ids] if len(pm.neuron_ids) == len(pm.neuron_indices) else [],
"selection_rule": pm.selection_rule,
"confidence": float(pm.confidence),
}
np.savez_compressed(
path,
neuron_ids=self.current_graph.neuron_ids,
coordinates=self.current_graph.coordinates,
tbars=self.current_graph.tbars,
sides=np.array(self.current_graph.sides),
row_offsets=self.current_graph.row_offsets,
col_indices=self.current_graph.col_indices,
weights=self.current_graph.weights,
mode=np.array(self.current_graph.mode.value),
generation=np.array(self.generation),
brain_id=np.array(self.current_brain_id),
seed=np.array(self.seed),
history_json=np.array(json.dumps(self.history)),
populations_json=np.array(json.dumps(pop_ser)),
)
self.save_history()
return path
@classmethod
def load_checkpoint(cls, checkpoint_file: str,
history_file: str = "diagnostics/evolution_history.json") -> "EvolutionScheduler":
"""Resumes a scheduler bit-exactly; continued trajectory matches uninterrupted run."""
from src.connectome.types import (GraphMode, MaleCNSRealGraph,
MaleCNSSpatialSurrogateGraph, SyntheticTestGraph,
PopulationMetadata, PopulationRegistry,
ProvenanceStatus)
data = np.load(checkpoint_file, allow_pickle=True)
mode = GraphMode(str(data["mode"]))
cls_map = {GraphMode.REAL: MaleCNSRealGraph,
GraphMode.SPATIAL_SURROGATE: MaleCNSSpatialSurrogateGraph,
GraphMode.SYNTHETIC_TEST: SyntheticTestGraph}
ids = np.array(data["neuron_ids"])
registry = PopulationRegistry()
pop_ser = json.loads(str(data["populations_json"]))
for name, ps in pop_ser.items():
idx = np.array(ps["indices"], dtype=np.int32)
registry.register(PopulationMetadata(
name=name, source="checkpoint-resume",
selection_rule=ps.get("selection_rule", ""),
neuron_ids=ids[idx] if len(idx) else np.array([], dtype=np.int64),
neuron_indices=idx, count=len(idx),
provenance_status=ProvenanceStatus.DERIVED,
confidence=float(ps.get("confidence", 0.9)),
))
graph = cls_map[mode](
neuron_ids=ids,
coordinates=np.array(data["coordinates"]),
tbars=np.array(data["tbars"]),
sides=list(data["sides"]),
row_offsets=np.array(data["row_offsets"]),
col_indices=np.array(data["col_indices"]),
weights=np.array(data["weights"]),
populations=registry,
)
sched = cls(base_graph=graph, history_file=history_file, seed=int(data["seed"]))
sched.generation = int(data["generation"])
sched.current_brain_id = str(data["brain_id"])
sched.history = json.loads(str(data["history_json"]))
return sched
|