凸优化
几何中心
给定一些二维平面上的点,求一个目标点使得其到所有给定点的欧氏距离之和最小。
\[
\displaylines{
\min_{x,y}\sum_{i}\sqrt{(x-x_i)^2 + (y-y_i)^2}
}
\]
没有解析解,但目标函数为凸函数,故使用梯度下降法求解:
class Solution {
public:
double getMinDistSum(vector<vector<int>>& positions) {
double eps = 1e-7;
double alpha = 1;
double decay = 1e-3;
int n = positions.size();
int batchSize = min(n, 32);
// mean position
double x = 0.0, y = 0.0;
for (const auto& pos: positions) {
x += pos[0];
y += pos[1];
}
x /= n;
y /= n;
// random engine
mt19937 gen{random_device{}()};
// loop epoch until convergence
while (true) {
shuffle(positions.begin(), positions.end(), gen);
double xPrev = x;
double yPrev = y;
// mini-batch SGD
for (int i = 0; i < n; i += batchSize) {
int j = min(i + batchSize, n);
double dx = 0.0, dy = 0.0;
// accumulate gradient per batch
for (int k = i; k < j; ++k) {
const auto& pos = positions[k];
dx += (x - pos[0]) / (sqrt((x - pos[0]) * (x - pos[0]) + (y - pos[1]) * (y - pos[1])) + eps);
dy += (y - pos[1]) / (sqrt((x - pos[0]) * (x - pos[0]) + (y - pos[1]) * (y - pos[1])) + eps);
}
// update
x -= alpha * dx;
y -= alpha * dy;
// lower lr
alpha *= (1.0 - decay);
}
// determine convergence
if (sqrt((x - xPrev) * (x - xPrev) + (y - yPrev) * (y - yPrev)) < eps) {
break;
}
}
// get objective function value
auto getDist = [&](double xc, double yc) {
double ans = 0;
for (const auto& pos: positions) {
ans += sqrt((pos[0] - xc) * (pos[0] - xc) + (pos[1] - yc) * (pos[1] - yc));
}
return ans;
};
return getDist(x, y);
}
};