#826586 (C++17) No.5010 Better Mo's Algorithm is Needed!! (Unweighted)

• 問題: No.5010 Better Mo's Algorithm is Needed!! (Unweighted)
• ユーザー: t33f
• 提出日時: 2022-12-17 23:08:51
• 言語: C++17 (gcc 12.3.0 + boost 1.83.0)
• 結果: AC
• 実行時間: 4,974 ms / 5,000 ms
• コード長: 6,019 bytes
• コンパイル時間: 2,420 ms
• 実行使用メモリ: 8,156 KB
• スコア: 19,839,777,113
• 最終ジャッジ日時: 2022-12-17 23:21:02
• 合計ジャッジ時間: 646,614 ms
• ジャッジサーバーID （参考情報）: judge15 / judge12

入力	結果	実行時間; 実行使用メモリ
testcase_00	AC	4,972 ms; 7,980 KB
testcase_01	AC	4,972 ms; 7,844 KB
testcase_02	AC	4,972 ms; 8,012 KB
testcase_03	AC	4,972 ms; 7,924 KB
testcase_04	AC	4,973 ms; 7,936 KB
testcase_05	AC	4,972 ms; 7,936 KB
testcase_06	AC	4,973 ms; 7,984 KB
testcase_07	AC	4,972 ms; 7,948 KB
testcase_08	AC	4,973 ms; 7,848 KB
testcase_09	AC	4,972 ms; 7,924 KB
testcase_10	AC	4,973 ms; 7,936 KB
testcase_11	AC	4,972 ms; 7,984 KB
testcase_12	AC	4,973 ms; 8,032 KB
testcase_13	AC	4,972 ms; 7,936 KB
testcase_14	AC	4,972 ms; 7,964 KB
testcase_15	AC	4,973 ms; 7,904 KB
testcase_16	AC	4,972 ms; 7,956 KB
testcase_17	AC	4,973 ms; 7,844 KB
testcase_18	AC	4,972 ms; 7,880 KB
testcase_19	AC	4,973 ms; 7,944 KB
testcase_20	AC	4,972 ms; 7,836 KB
testcase_21	AC	4,973 ms; 7,936 KB
testcase_22	AC	4,972 ms; 7,924 KB
testcase_23	AC	4,972 ms; 7,880 KB
testcase_24	AC	4,972 ms; 7,980 KB
testcase_25	AC	4,971 ms; 7,916 KB
testcase_26	AC	4,972 ms; 7,900 KB
testcase_27	AC	4,971 ms; 7,976 KB
testcase_28	AC	4,972 ms; 7,980 KB
testcase_29	AC	4,972 ms; 7,852 KB
testcase_30	AC	4,971 ms; 7,884 KB
testcase_31	AC	4,973 ms; 7,872 KB
testcase_32	AC	4,972 ms; 7,904 KB
testcase_33	AC	4,973 ms; 7,908 KB
testcase_34	AC	4,972 ms; 7,976 KB
testcase_35	AC	4,973 ms; 7,912 KB
testcase_36	AC	4,972 ms; 7,916 KB
testcase_37	AC	4,974 ms; 7,848 KB
testcase_38	AC	4,972 ms; 7,860 KB
testcase_39	AC	4,972 ms; 7,916 KB
testcase_40	AC	4,973 ms; 7,856 KB
testcase_41	AC	4,972 ms; 7,992 KB
testcase_42	AC	4,973 ms; 7,932 KB
testcase_43	AC	4,972 ms; 7,908 KB
testcase_44	AC	4,973 ms; 7,888 KB
testcase_45	AC	4,972 ms; 7,880 KB
testcase_46	AC	4,973 ms; 7,964 KB
testcase_47	AC	4,972 ms; 7,908 KB
testcase_48	AC	4,973 ms; 7,968 KB
testcase_49	AC	4,972 ms; 7,844 KB
testcase_50	AC	4,972 ms; 7,916 KB
testcase_51	AC	4,972 ms; 7,844 KB
testcase_52	AC	4,972 ms; 8,008 KB
testcase_53	AC	4,972 ms; 7,852 KB
testcase_54	AC	4,972 ms; 7,884 KB
testcase_55	AC	4,973 ms; 7,924 KB
testcase_56	AC	4,972 ms; 7,880 KB
testcase_57	AC	4,973 ms; 7,876 KB
testcase_58	AC	4,973 ms; 7,916 KB
testcase_59	AC	4,972 ms; 7,856 KB
testcase_60	AC	4,973 ms; 8,020 KB
testcase_61	AC	4,972 ms; 7,900 KB
testcase_62	AC	4,973 ms; 7,848 KB
testcase_63	AC	4,972 ms; 7,884 KB
testcase_64	AC	4,973 ms; 7,916 KB
testcase_65	AC	4,973 ms; 7,932 KB
testcase_66	AC	4,972 ms; 7,908 KB
testcase_67	AC	4,973 ms; 7,952 KB
testcase_68	AC	4,972 ms; 7,852 KB
testcase_69	AC	4,973 ms; 7,848 KB
testcase_70	AC	4,972 ms; 7,844 KB
testcase_71	AC	4,973 ms; 7,988 KB
testcase_72	AC	4,972 ms; 7,936 KB
testcase_73	AC	4,973 ms; 7,940 KB
testcase_74	AC	4,972 ms; 7,848 KB
testcase_75	AC	4,973 ms; 7,956 KB
testcase_76	AC	4,972 ms; 7,840 KB
testcase_77	AC	4,972 ms; 7,932 KB
testcase_78	AC	4,973 ms; 7,848 KB
testcase_79	AC	4,972 ms; 7,952 KB
testcase_80	AC	4,973 ms; 7,840 KB
testcase_81	AC	4,973 ms; 8,008 KB
testcase_82	AC	4,973 ms; 7,932 KB
testcase_83	AC	4,972 ms; 7,940 KB
testcase_84	AC	4,973 ms; 7,912 KB
testcase_85	AC	4,972 ms; 7,964 KB
testcase_86	AC	4,973 ms; 7,960 KB
testcase_87	AC	4,972 ms; 7,864 KB
testcase_88	AC	4,972 ms; 7,996 KB
testcase_89	AC	4,973 ms; 8,156 KB
testcase_90	AC	4,972 ms; 7,964 KB
testcase_91	AC	4,972 ms; 7,904 KB
testcase_92	AC	4,972 ms; 7,972 KB
testcase_93	AC	4,972 ms; 7,832 KB
testcase_94	AC	4,973 ms; 7,964 KB
testcase_95	AC	4,973 ms; 7,952 KB
testcase_96	AC	4,973 ms; 7,924 KB
testcase_97	AC	4,972 ms; 7,840 KB
testcase_98	AC	4,973 ms; 8,032 KB
testcase_99	AC	4,972 ms; 7,840 KB
testcase_100	AC	4,973 ms; 7,916 KB
testcase_101	AC	4,972 ms; 7,868 KB
testcase_102	AC	4,973 ms; 8,052 KB
testcase_103	AC	4,972 ms; 7,928 KB
testcase_104	AC	4,972 ms; 7,852 KB
testcase_105	AC	4,973 ms; 7,932 KB
testcase_106	AC	4,972 ms; 8,020 KB
testcase_107	AC	4,973 ms; 7,976 KB
testcase_108	AC	4,972 ms; 7,876 KB
testcase_109	AC	4,973 ms; 7,932 KB
testcase_110	AC	4,972 ms; 7,844 KB
testcase_111	AC	4,973 ms; 7,924 KB
testcase_112	AC	4,972 ms; 7,936 KB
testcase_113	AC	4,973 ms; 7,976 KB
testcase_114	AC	4,974 ms; 7,916 KB
testcase_115	AC	4,972 ms; 7,944 KB
testcase_116	AC	4,973 ms; 7,924 KB
testcase_117	AC	4,973 ms; 7,880 KB
testcase_118	AC	4,972 ms; 7,916 KB
testcase_119	AC	4,972 ms; 7,936 KB

ソースコード

#include <cstdlib>
#include <cmath>
#include <algorithm>
#include <chrono>
#include <cassert>
#include <array>
#include <vector>
#include <numeric>
#include <cstdio>
#include <iostream>
using namespace std;

unsigned int xor128() {
    static unsigned int x=123456789, y=362436069, z=521288629, w=88675123;
    unsigned int t;
    t=(x^(x<<11)); x=y; y=z; z=w;
    return (w=(w^(w>>19))^(t^(t>>8)));
}

inline bool rand_bool(double prob) {
    constexpr double x = 1LL<<32; // uint_max+1
    return xor128() < prob * x;
}

int timelimit = 5 * 1000, margin = 50;

class Timer {
  chrono::system_clock::time_point start_time = chrono::system_clock::now();
public:
  Timer() {}
  long long get_elapsed_time() {
    auto diff = chrono::system_clock::now() - start_time;
    return chrono::duration_cast<chrono::milliseconds>(diff).count();
  }
} timer;


constexpr int N = 200000, Q = 200000;
int WT, ST;
int L[Q], R[Q];

long long initial_cost(int i) {
    return 100 * (R[i] - L[i] + 1);
}

long long adj_cost(int i, int j) {
    return 100 * (abs(L[i] - L[j]) + abs(R[i] - R[j]));
}

long long calc_cost(const array<int, Q> &ans) {
    long long c = initial_cost(ans[0]);
    for (int i = 1; i < Q; ++i) c += adj_cost(ans[i - 1], ans[i]);
    return c;
}

long long score_of_cost(long long cost) {
    constexpr long long n = (long long)1e18;
    return n % cost * 2 >= cost ? n / cost + 1 : n / cost;
}

void read_input() {
    scanf("%*d %*d %d %d", &WT, &ST);
    assert(WT == 1);
    for (int i = 0; i < N; i++) scanf(" %*d");
    for (int i = 0; i < Q; i++) scanf(" %d %d", L + i, R + i);
}

vector<vector<int>> partition_queries(int B) {
    if (ST != 1) {
        array<int, N> start_count;
        for (int i = 0; i < Q; ++i) start_count[L[i]]++;

        vector<int> boundaries;
        long long w1 = 0, w2 = Q;
        long long s = 0;
        const int bias = 40;
        for (int i = 0; i < N; ++i) {
            s += start_count[i];
            w1 += s;
            w2 -= 1;
            if (w1 > w2 * bias) {
                boundaries.push_back(i);
                s = 0;
                w1 = 0;
            }
        }

        vector<vector<int>> buckets(boundaries.size() + 1);
        boundaries.push_back(N); // sentinel
        for (int i = 0; i < Q; ++i) {
            auto it = lower_bound(boundaries.begin(), boundaries.end(), L[i]);
            const size_t j = it - boundaries.begin();
            buckets[j].push_back(i);
        }

        return buckets;
    } else {
        array<int, N> a;
        iota(a.begin(), a.end(), 0);
        vector<vector<int>> buckets((N + B - 1) / B + 1);
        for (int i = 0; i < Q; i++) {
            const int b = L[i] / B;
            buckets[b].push_back(i);
        }
        return buckets;
    }
}

array<int, Q> initial_solution(int B) {
    vector<vector<int>> buckets = partition_queries(B);
    array<int, Q> ans;
    int pos = 0;
    for (int b = 0; b < buckets.size(); ++b) {
        vector<int> &elems = buckets[b];
        sort(elems.begin(), elems.end(), [b](int i, int j) {
            if (b % 2 == 0) return make_pair(R[i], L[i]) < make_pair(R[j], L[j]);
            else return make_pair(R[i], L[i]) > make_pair(R[j], L[j]);
        });
        for (int i : elems) ans[pos++] = i;
    }
    return ans;
}

array<int, Q> solve(int B, int S) {
    array<int, Q> ans = initial_solution(B);
    long long cur_cost = calc_cost(ans);
    long long best_cost = cur_cost;
    array<int, Q> best_ans = ans;
    const double start_temp = cur_cost / N * 0.15, end_temp = 0;
    const int upd_interval = 10;
    long long next_upd = timelimit - margin - 10 * upd_interval;
    for (;;) {
        cerr << cur_cost << ' ' << score_of_cost(cur_cost) << endl;
        for (int i = 0; i < N; ++i) {
            const long long elapsed = timer.get_elapsed_time();
            const bool finish = (elapsed + margin > timelimit);
            const bool check_update = (finish || next_upd < elapsed);
            // best_ans の更新がボトルネックになるので毎回更新しないで
            // 最後の方だけ一定時間ごとに現在のスコアを見て、改善していたら更新
            if (check_update && cur_cost < best_cost) {
                next_upd += upd_interval;
                best_cost = cur_cost;
                best_ans = ans;
            }
            if (finish) {
                if (initial_cost(best_ans.front()) > initial_cost(best_ans.back()))
                    reverse(best_ans.begin(), best_ans.end());
                return best_ans;
            }

            const double progress = double(elapsed) / timelimit,
                temp = start_temp + (end_temp - start_temp) * progress;
            const int jmax = min(N, i + S);
            for (int j = i + 1; j < jmax; ++j) {
                long long diff = 0;
                if (i > 0) {
                    diff -= adj_cost(ans[i - 1], ans[i]);
                    diff += adj_cost(ans[i - 1], ans[j]);
                } else {
                    diff -= initial_cost(ans[0]);
                    diff += initial_cost(ans[j]);
                }
                if (j + 1 < N) {
                    diff -= adj_cost(ans[j], ans[j + 1]);
                    diff += adj_cost(ans[i], ans[j + 1]);
                }
                bool accept = (diff <= 0);
                if (!accept && diff < temp * 3) {
                    double prob = exp(-diff / temp);
                    accept = rand_bool(prob);
                }
                if (accept) {
                    reverse(ans.begin() + i, ans.begin() + j + 1);
                    cur_cost += diff;
                }
            }
        }
    }
}

int main(int argc, char **argv) {
    read_input();

    const int B = argc <= 1 ? 3500 : atoi(argv[1]);
    const int S = argc <= 2 ? 100 : atoi(argv[2]);
    const array<int, Q> ans = solve(B, S);

    for (int i : ans) cout << i + 1 << ' ';
    cout << endl;

    cerr << "score = " << score_of_cost(calc_cost(ans)) << endl;
}