File size: 6,151 Bytes
3a3eea5 | 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 | #!/usr/bin/env python3
"""Stress tests for graph misspecification, workflow size, and verifier quality."""
from __future__ import annotations
import argparse
import csv
import math
import random
from pathlib import Path
from statistics import mean, stdev
from run_experiments import Graph, choose, episode, make_graph
KINDS = ["chain", "tree", "diamond", "random"]
def ci95(values: list[float]) -> float:
return 1.96 * stdev(values) / math.sqrt(len(values))
def write_csv(path: Path, rows: list[dict]) -> None:
with path.open("w", newline="", encoding="utf-8") as handle:
writer = csv.DictWriter(handle, fieldnames=list(rows[0]))
writer.writeheader()
writer.writerows(rows)
def perturb_edges(graph: Graph, drop_rate: float, add_rate: float, rng: random.Random) -> Graph:
n = len(graph.p)
true_edges = {(u, v): q for u, children in enumerate(graph.children) for v, q in children}
kept = {(u, v): q for (u, v), q in true_edges.items() if rng.random() >= drop_rate}
target_additions = round(len(true_edges) * add_rate)
candidates = [(u, v) for u in range(n) for v in range(u + 1, n)
if (u, v) not in true_edges]
rng.shuffle(candidates)
for u, v in candidates[:target_additions]:
kept[(u, v)] = rng.uniform(0.42, 0.88)
parents = [[] for _ in range(n)]
children = [[] for _ in range(n)]
for (u, v), q in sorted(kept.items()):
children[u].append((v, q))
parents[v].append((u, q))
# Terminal outputs are part of the workflow specification and remain known;
# only dependency edges are perturbed.
return Graph(parents=parents, children=children, p=list(graph.p), sinks=list(graph.sinks))
def evaluate(graph: Graph, selected: set[int], sensitivity: float,
episodes: int, seed_base: int) -> float:
return mean([
episode(graph, selected, sensitivity, random.Random(seed_base + ep))[0]
for ep in range(episodes)
])
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--out", type=Path, default=Path("results/structure_sensitivity"))
parser.add_argument("--graphs", type=int, default=240)
parser.add_argument("--episodes", type=int, default=400)
parser.add_argument("--seed", type=int, default=7)
args = parser.parse_args()
args.out.mkdir(parents=True, exist_ok=True)
master = random.Random(args.seed)
graphs = [(kind, make_graph(kind, 40, random.Random(master.randrange(2**63))))
for kind in KINDS for _ in range(args.graphs // len(KINDS))]
topology_rows: list[dict] = []
conditions = [
("exact", 0.0, 0.0),
("drop", 0.1, 0.0), ("drop", 0.2, 0.0), ("drop", 0.3, 0.0), ("drop", 0.5, 0.0),
("add", 0.0, 0.1), ("add", 0.0, 0.2), ("add", 0.0, 0.3), ("add", 0.0, 0.5),
("mixed", 0.1, 0.1), ("mixed", 0.2, 0.2), ("mixed", 0.3, 0.3),
]
for label, drop, add in conditions:
values = {"local-risk": [], "toporisk": []}
for graph_id, (_, graph) in enumerate(graphs):
estimated = perturb_edges(
graph, drop, add,
random.Random(args.seed * 1_000_003 + graph_id * 101 + int(drop * 100) * 3 + int(add * 100)),
)
for method_id, method in enumerate(values):
selected = choose(estimated, method, 4, random.Random(graph_id + method_id))
values[method].append(evaluate(
graph, selected, .90, args.episodes,
args.seed * 10**8 + graph_id * 10_000,
))
for method, observed in values.items():
topology_rows.append({
"condition": label,
"drop_rate": drop,
"add_rate": add,
"method": method,
"terminal_error": mean(observed),
"terminal_error_ci95": ci95(observed),
})
write_csv(args.out / "topology_misspecification.csv", topology_rows)
sensitivity_rows: list[dict] = []
for sensitivity in [.50, .70, .90, 1.00]:
values = {"local-risk": [], "toporisk": []}
for graph_id, (_, graph) in enumerate(graphs):
for method_id, method in enumerate(values):
selected = choose(graph, method, 4, random.Random(graph_id + method_id))
values[method].append(evaluate(
graph, selected, sensitivity, args.episodes,
args.seed * 10**8 + graph_id * 10_000,
))
for method, observed in values.items():
sensitivity_rows.append({
"sensitivity": sensitivity,
"method": method,
"terminal_error": mean(observed),
"terminal_error_ci95": ci95(observed),
})
write_csv(args.out / "verifier_sensitivity.csv", sensitivity_rows)
size_rows: list[dict] = []
size_master = random.Random(args.seed + 884_221)
for n in [20, 40, 80, 160]:
size_graphs = [(kind, make_graph(kind, n, random.Random(size_master.randrange(2**63))))
for kind in KINDS for _ in range(30)]
values = {"local-risk": [], "toporisk": []}
for graph_id, (_, graph) in enumerate(size_graphs):
k = max(1, round(.10 * n))
for method_id, method in enumerate(values):
selected = choose(graph, method, k, random.Random(graph_id + method_id))
values[method].append(evaluate(
graph, selected, .90, 200,
args.seed * 10**8 + n * 1_000_000 + graph_id * 1_000,
))
for method, observed in values.items():
size_rows.append({
"nodes": n,
"graphs": len(size_graphs),
"episodes_per_graph": 200,
"method": method,
"terminal_error": mean(observed),
"terminal_error_ci95": ci95(observed),
})
write_csv(args.out / "workflow_size.csv", size_rows)
print(f"wrote three stress-test tables to {args.out}")
if __name__ == "__main__":
main()
|