shun-quantum-icc/add_linkselfie/algorithms/lonline_nb.py

import math

def lonline_network_benchmarking(network, path_list, bounces, C_budget):
    candidate_set = list(path_list)
    s = 0
    C = 0.01     
    delta = 0.1  
    cost = 0
    estimated_fidelities = {}

    # 1経路を1サンプル測る想定コスト（list 前提）
    cost_per_sample_unit = sum(bounces) if sum(bounces) > 0 else 1

    while cost < C_budget and len(candidate_set) > 1:
        s += 1
        Ns = math.ceil(C * (2 ** (2 * s)) * math.log2(max((2 ** s) * len(candidate_set) / delta, 2)))
        if Ns < 4:
            Ns = 4

        # --- 事前コスト見積り（lonline準拠） ---
        cost_needed_for_one_path = Ns * cost_per_sample_unit
        # 2ラウンド目以降で1経路ぶんすら入らないなら終了
        if cost + cost_needed_for_one_path > C_budget and s > 1:
            break
        # 1ラウンド目だけは Ns を縮退してでも1回は回す（入らなければ中止）
        """
        if cost + cost_needed_for_one_path > C_budget and s == 1:
            Ns_fit = (C_budget - cost) // max(cost_per_sample_unit, 1)
            if Ns_fit <= 0:
                break
            Ns = int(Ns_fit)
            cost_needed_for_one_path = Ns * cost_per_sample_unit
        """
        # ---------------------------------------

        sample_times = {i: Ns for i in bounces}

        p_s = {}
        measured_paths = []  # このラウンドで実際に測れた経路だけを削除判定に使う
        for path in list(candidate_set):
            if cost + cost_needed_for_one_path > C_budget:
                continue  # 予算に入らない経路はこのラウンドはスキップ
            p, bounces_num = network.benchmark_path(path, bounces, sample_times)
            cost += bounces_num
            estimated_fidelities[path] = p + (1 - p) / 2  # ★ 既存式＆変数名を踏襲
            p_s[path] = p
            measured_paths.append(path)

        if not p_s:
            break  # このラウンドで1つも測れなかった

        # online_nb.py と同じ 2^{-s} 幅の連続削除
        p_max = max(p_s.values())
        new_candidate_set = []
        for path in measured_paths:  # 測れたものだけで判定（KeyError防止）
            if p_s[path] + 2**(-s) > p_max - 2**(-s):
                new_candidate_set.append(path)

        # もし全消しになったら、保険として現集合を維持
        candidate_set = new_candidate_set or candidate_set

    if not estimated_fidelities:
        return None, cost, None

    best_path = max(estimated_fidelities, key=estimated_fidelities.get)
    best_path_fidelity = estimated_fidelities[best_path]
    correctness = (best_path == getattr(network, "best_path", None))
    return correctness, cost, best_path_fidelity