| 12345678910111213141516171819202122232425262728293031323334353637383940 |
- # lnaive_nb.py
- def naive_network_benchmarking_with_budget(network, path_list, bounces, C_budget):
- """
- 目的:
- 総予算 C_budget を各パスへ均等割り当てし、均等サンプリングで NB を実行。
- 実行コストは常に予算内(超過しない)。
- 出力:
- (correctness, cost, best_path_fidelity)
- correctness … 推定最良パスが真の最良と一致したか
- cost … 実測で消費した総コスト
- best_path_fidelity … 推定最良パスの推定忠実度(naive変換後)
- """
- fidelity, cost = {}, 0
- n_paths = len(path_list)
- if n_paths == 0:
- return False, 0, None
- per_sample_cost = sum(bounces) or 1
- per_path_budget = int(C_budget) // n_paths
- Ns = per_path_budget // per_sample_cost # 各パスのサンプル数
- if Ns <= 0:
- return False, 0, None
- # 各 hop に同じ Ns を配る(既存 naive と同じ割当表)
- sample_times = {h: int(Ns) for h in bounces}
- # 各パスを均等回数でベンチマーク
- for path in path_list:
- p, used = network.benchmark_path(path, bounces, sample_times)
- fidelity[path] = p + (1 - p) / 2 # 既存 naive と同じ忠実度変換
- cost += used
- if not fidelity:
- return False, cost, None
- best_path = max(fidelity, key=fidelity.get)
- correctness = (best_path == getattr(network, "best_path", None))
- best_path_fidelity = fidelity[best_path]
- return correctness, cost, best_path_fidelity
|