八数码问题(8-Puzzle Problem)——多种搜索算法
八数码问题(8-Puzzle Problem)——多种搜索算法
P1379 八数码难题 - 洛谷
题目概述
在 \(3 \times 3\) 的棋盘上摆放着 \(8\) 个棋子,棋子的编号分别为 \(1\) 到 \(8\),空格则用 \(0\) 表示。与空格直接相连的棋子可以移至空格中,这样原来棋子的位置就成为空格。现给出一种初始布局,求到达目标布局的最少步数。为简单起见,目标布局总是如下:
123 804 765
本题是一道经典的搜索题,下面将介绍几种常见的搜索算法。以下所有代码均需要 C++11 标准。
朴素 BFS
通过对题目的简单分析,很容易写出朴素的 BFS 代码。进行访问标记时,可以利用哈希的思想,将矩阵转化为整数,再用 std::unordered_set
存储。由于本题的数据范围较小,朴素的 BFS 算法也能通过本题测试,但是效率较低。具体代码如下:
View Code
#include <bits/stdc++.h>
using namespace std;
const int tar_x = 2, tar_y = 2, target = 123804765;
const int dx[] = { 1, -1, 0, 0 };
const int dy[] = { 0, 0, 1, -1 };
struct Status {
int maze[5][5]; // matrix
int x, y; // coordinate of blank space
int t; // step number
explicit Status(int num) {
memset(maze, 0, sizeof(maze));
t = 0;
for (int i = 3; i >= 1; --i) {
for (int j = 3; j >= 1; --j) {
maze[i][j] = num % 10;
num /= 10;
if (maze[i][j] == 0)
x = i, y = j;
}
}
}
int to_int() const {
int ans = 0;
for (int i = 1; i <= 3; ++i)
for (int j = 1; j <= 3; ++j)
ans = ans * 10 + maze[i][j]; // hash
return ans;
}
};
int bfs(int num) {
queue<Status> q;
unordered_set<int> vis;
q.emplace(num);
vis.insert(num);
while (!q.empty()) {
Status now = q.front();
q.pop();
if (now.x == tar_x && now.y == tar_y && now.to_int() == target)
return now.t; // exit
++now.t;
int x = now.x, y = now.y;
for (int i = 0; i < 4; ++i) {
int tx = x + dx[i], ty = y + dy[i];
if (tx < 1 || tx > 3 || ty < 1 || ty > 3)
continue;
swap(now.maze[x][y], now.maze[tx][ty]);
now.x = tx;
now.y = ty;
if (!vis.count(now.to_int())) {
q.push(now); // expand
vis.insert(now.to_int());
}
now.x = x;
now.y = y;
swap(now.maze[x][y], now.maze[tx][ty]); // backtrack
}
}
return -1; // unused value
}
int main() {
int num;
cin >> num;
cout << bfs(num) << endl;
return 0;
}
双向 BFS
对于本题这类已知初始状态和目标状态的题目,可以考虑双向 BFS。在搜索开始前,同时将初始状态和目标状态放进 BFS 队列中。搜索过程中,标记每个状态被访问时的搜索方向以及从对应起点出发的步数。当一种状态被两个方向同时搜到,也就是两个方向相遇时,这两个方向的步数之和就是所求答案。BFS 的性质保证了这一答案一定是最小值。这样的算法称为 Meet in the Middle,通过将实际拓展的层数减半,大大提高了搜索效率,避免了许多不必要的状态拓展。具体代码如下:
View Code
#include <bits/stdc++.h>
using namespace std;
const int target = 123804765;
const int dx[] = { 1, -1, 0, 0 };
const int dy[] = { 0, 0, 1, -1 };
struct Status {
int maze[5][5]; // matrix
int x, y; // coordinate of blank space
bool d; // bfs direction (true: forward, false: back)
int t; // step number
explicit Status(int num) {
memset(maze, 0, sizeof(maze));
t = 0;
if (num == target)
d = false;
else
d = true;
for (int i = 3; i >= 1; --i) {
for (int j = 3; j >= 1; --j) {
maze[i][j] = num % 10;
num /= 10;
if (maze[i][j] == 0)
x = i, y = j;
}
}
}
int to_int() const {
int ans = 0;
for (int i = 1; i <= 3; ++i)
for (int j = 1; j <= 3; ++j)
ans = ans * 10 + maze[i][j]; // hash
return ans;
}
};
int bfs(int num) {
queue<Status> q;
unordered_map<int, pair<int, bool>> vis;
q.emplace(target); // target state
vis[target] = make_pair(0, false);
q.emplace(num); // starting state
vis[num] = make_pair(0, true);
while (!q.empty()) {
Status now = q.front();
q.pop();
if (vis.count(now.to_int()) && vis[now.to_int()].second != now.d)
return now.t + vis[now.to_int()].first; // meet in the middle
++now.t;
int x = now.x, y = now.y;
for (int i = 0; i < 4; ++i) {
int tx = x + dx[i], ty = y + dy[i];
if (tx < 1 || tx > 3 || ty < 1 || ty > 3)
continue;
swap(now.maze[x][y], now.maze[tx][ty]);
now.x = tx;
now.y = ty;
if (!vis.count(now.to_int()) || vis[now.to_int()].second != now.d) {
q.push(now); // expand
vis[now.to_int()] = make_pair(now.t, now.d);
}
now.x = x;
now.y = y;
swap(now.maze[now.x][now.y], now.maze[tx][ty]); // backtrack
}
}
return -1; // unused value
}
int main() {
int num;
cin >> num;
cout << bfs(num) << endl;
return 0;
}
A*
A* 算法是一种启发式搜索,即利用估值函数进行剪枝,以避免盲目搜索中许多不必要的状态拓展。A* 算法以 BFS 为基础,用优先队列代替 BFS 队列,以估值函数为优先级。A* 算法中,每个状态的估值函数由两部分组成,即 \(f(x)=g(x)+h(x)\),其中 \(g(x)\) 是已经走过的步数,\(h(x)\) 是预估到达终点至少还要走的步数,两者之和 \(f(x)\) 即这一状态的估值函数。因此,为确保算法正确,\(h(x)\) 的值一定不大于实际距离终点的步数,即 \(f(x)\) 的值一定不大于实际总步数。本题中,可以使用每个棋子到目标位置的曼哈顿距离作为其 \(h(x)\)。容易证明,该函数满足上述条件。具体代码如下:
View Code
#include <bits/stdc++.h>
using namespace std;
const int dx[] = { 1, -1, 0, 0 };
const int dy[] = { 0, 0, 1, -1 };
const int pos_x[] = { 2, 1, 1, 1, 2, 3, 3, 3, 2 };
const int pos_y[] = { 2, 1, 2, 3, 3, 3, 2, 1, 1 };
struct Status {
int maze[5][5]; // matrix
int x, y; // coordinate of blank space
int t; // step number
explicit Status(int num) {
memset(maze, 0, sizeof(maze));
t = 0;
for (int i = 3; i >= 1; --i) {
for (int j = 3; j >= 1; --j) {
maze[i][j] = num % 10;
num /= 10;
if (maze[i][j] == 0)
x = i, y = j;
}
}
}
int h() const {
int ans = 0;
for (int i = 1; i <= 3; ++i)
for (int j = 1; j <= 3; ++j)
if (maze[i][j] != 0)
ans += abs(i - pos_x[maze[i][j]]) + abs(j - pos_y[maze[i][j]]); // Manhattan distance
return ans;
}
int to_int() const {
int ans = 0;
for (int i = 1; i <= 3; ++i)
for (int j = 1; j <= 3; ++j)
ans = ans * 10 + maze[i][j]; // hash
return ans;
}
bool operator<(const Status& other) const {
return h() + t > other.h() + other.t; // compare by f(x)
}
};
int a_star(int num) {
priority_queue<Status, vector<Status>> pq;
set<int> vis;
pq.push(Status(num));
vis.insert(num);
while (!pq.empty()) {
if (pq.top().h() == 0)
return pq.top().t; // exit
Status now = pq.top();
pq.pop();
++now.t;
int x = now.x, y = now.y;
for (int i = 0; i < 4; ++i) {
int tx = x + dx[i], ty = y + dy[i];
if (tx < 1 || tx > 3 || ty < 1 || ty > 3)
continue;
swap(now.maze[now.x][now.y], now.maze[tx][ty]);
now.x = tx;
now.y = ty;
if (!vis.count(now.to_int())) {
pq.push(now); // expand
vis.insert(now.to_int());
}
now.x = x;
now.y = y;
swap(now.maze[now.x][now.y], now.maze[tx][ty]); // backtrack
}
}
return -1; // unused value
}
int main() {
int num;
cin >> num;
cout << a_star(num) << endl;
return 0;
}
IDA*
IDA* 就是基于迭代加深搜索的 A* 算法。所谓迭代加深,就是在 DFS 的基础上控制其搜索深度,一旦超过深度限制就停止搜索,若当前深度无法得到答案,则再增加深度限制。迭代加深搜索结合了 DFS 与 BFS 的优点,不需要占用大量空间,支持回溯,同时可以快速找到最优解,避免剪枝不充分而造成的大量无用搜素,并且不需要判重。此外,由于迭代加深算法基于 DFS,相对于 BFS 而言,其实现难度更低,代码量更少。IDA* 则是在迭代加深搜素的基础上加上了估值函数的剪枝。有关估值函数的内容,在 A* 部分 已经说明,此处不再赘述。具体代码如下:
View Code
#include <bits/stdc++.h>
using namespace std;
const int dx[] = { 1, -1, 0, 0 };
const int dy[] = { 0, 0, 1, -1 };
const int pos_x[] = { 2, 1, 1, 1, 2, 3, 3, 3, 2 };
const int pos_y[] = { 2, 1, 2, 3, 3, 3, 2, 1, 1 };
int lim; // depth limit
int m[5][5];
int h() {
int ans = 0;
for (int i = 1; i <= 3; ++i)
for (int j = 1; j <= 3; ++j)
if (m[i][j] != 0)
ans += abs(i - pos_x[m[i][j]]) + abs(j - pos_y[m[i][j]]); // Manhattan distance
return ans;
}
bool dfs(int x, int y, int t, int lx, int ly) {
int dis = h();
if (t + dis > lim)
return false; // prune with f(x)
if (dis == 0)
return true; // exit
for (int i = 0; i < 4; ++i) {
int tx = x + dx[i], ty = y + dy[i];
if (tx < 1 || tx > 3 || ty < 1 || ty > 3)
continue;
if (tx == lx && ty == ly)
continue; // very important
swap(m[x][y], m[tx][ty]);
if (dfs(tx, ty, t + 1, x, y))
return true; // expand
swap(m[x][y], m[tx][ty]); // backtrack
}
return false;
}
int main() {
int num;
cin >> num;
int sx, sy;
for (int i = 3; i >= 1; --i) {
for (int j = 3; j >= 1; --j) {
m[i][j] = num % 10;
num /= 10;
if (m[i][j] == 0)
sx = i, sy = j;
}
}
lim = 0;
while (!dfs(sx, sy, 0, -1, -1))
++lim; // IDA*
cout << lim << endl;
return 0;
}
转载请注明出处。原文地址:https://www.cnblogs.com/na-sr/p/8-puzzle.html