OSQP官方文档
1 QSQP简介
OSQP求解形式为的凸二次规划:
x
∈
R
n
x∈R^n
x∈Rn:优化变量
P
∈
S
+
n
P∈S^n_+
P∈S+n:半正定矩阵
特征
(1)高效:使用了一种自定义的基于ADMM的一阶方法,只需要在设置阶段进行单个矩阵分解。
(2)鲁棒:该算法设置之后不需要对问题数据进行假设(问题只需要是凸的)。
(3)原始/对偶不可行问题:当问题是原始或对偶不可行时,OSQP会检测到它。这是第一个基于一阶方法的QP求解器。
(4)可嵌入:有一个简单的接口来生成定制的可嵌入C代码,而不需要内存管理器。
(5)不需要外部库即可运行
(6)可以很容易地进行热启动,并且可以缓存矩阵分解,以非常有效地解决参数化问题
(7)接口:提供了到C、C++、Fortran、Julia、Matlab、Python、R、Ruby和Rust的接口
2 OSQP求解器
求解器运行以下ADMM算法:
Π
\Pi
Π:投影到超盒上
[
l
,
u
]
[l,u]
[l,u],
ρ
\rho
ρ是ADMM步长
Linear system solution
线性系统解是算法的核心部分。它可以使用直接或间接的方法来完成。
使用直接线性系统求解器,我们求解以下具有拟定矩阵的线性系统:
使用间接线性系统求解器,我们求解以下具有正定矩阵的线性系统:
OSQP核心旨在支持不同的线性系统求解器。
Convergence
在每k次迭代时,OSQP生成一个元组
(
x
k
,
z
k
,
y
k
)
(x^k,z^k,y^k)
(xk,zk,yk),
x
k
∈
R
n
,
z
k
、
y
k
∈
R
m
x^k∈R^n,z^k、y^k∈R^m
xk∈Rn,zk、yk∈Rm
与
(
x
k
,
z
k
,
y
k
)
(x^k,z^k,y^k)
(xk,zk,yk)相关的原始残差和对偶残差:
不可行问题
OSQP能够检测问题是原始不可行还是对偶不可行。
3 Get started
安装:
Linux操作系统,默认gcc,cmake已经安装好
① 克隆存储库
git clone https://github.com/osqp/osqp
② 创建目录和更改目录build
cd osqp
mkdir build
cd build
③ 创建 Makefile
cmake -G "Unix Makefiles" ..
④ 编译 OSQP
cmake --build .
C语言:
在 CMake 项目中包括 OSQP,具体取决于您需要共享库还是静态库:
# Find OSQP library and headers
find_package(osqp REQUIRED)
# Link the OSQP shared library
target_link_libraries(yourTarget PRIVATE osqp::osqp)
# or...
# Link the OSQP static library
target_link_libraries(yourTarget PRIVATE osqp::osqpstatic)
4 接口
OSQP有几个接口。以下链接中显示了有关设置、状态值以及如何指定不同线性系统解算器的信息
Solver settings
Linear Systems Solvers
Status values
C:github.com/osqp/osqp
C++:github.com/robotology/osqp-eigen
5 Examples
Demo:
① Setup and solve :设置和求解
C:
#include <stdlib.h>
#include "osqp.h"
int main(int argc, char **argv) {
/* Load problem data */
OSQPFloat P_x[3] = {4.0, 1.0, 2.0, };
OSQPInt P_nnz = 3;
OSQPInt P_i[3] = {0, 0, 1, };
OSQPInt P_p[3] = {0, 1, 3, };
OSQPFloat q[2] = {1.0, 1.0, };
OSQPFloat A_x[4] = {1.0, 1.0, 1.0, 1.0, };
OSQPInt A_nnz = 4;
OSQPInt A_i[4] = {0, 1, 0, 2, };
OSQPInt A_p[3] = {0, 2, 4, };
OSQPFloat l[3] = {1.0, 0.0, 0.0, };
OSQPFloat u[3] = {1.0, 0.7, 0.7, };
OSQPInt n = 2;
OSQPInt m = 3;
/* Exitflag */
OSQPInt exitflag = 0;
/* Solver, settings, matrices */
OSQPSolver *solver;
OSQPSettings *settings;
OSQPCscMatrix* P = malloc(sizeof(OSQPCscMatrix));
OSQPCscMatrix* A = malloc(sizeof(OSQPCscMatrix));
/* Populate matrices */
csc_set_data(A, m, n, A_nnz, A_x, A_i, A_p);
csc_set_data(P, n, n, P_nnz, P_x, P_i, P_p);
/* Set default settings */
settings = (OSQPSettings *)malloc(sizeof(OSQPSettings));
if (settings) {
osqp_set_default_settings(settings);
settings->alpha = 1.0; /* Change alpha parameter */
}
/* Setup solver */
exitflag = osqp_setup(&solver, P, q, A, l, u, m, n, settings);
/* Solve problem */
if (!exitflag) exitflag = osqp_solve(solver);
/* Cleanup */
osqp_cleanup(solver);
if (A) free(A);
if (P) free(P);
if (settings) free(settings);
return (int)exitflag;
};
② Update vectors:更新向量
#include <stdlib.h>
#include "osqp.h"
int main(int argc, char **argv) {
/* Load problem data */
OSQPFloat P_x[3] = {4.0, 1.0, 2.0, };
OSQPInt P_nnz = 3;
OSQPInt P_i[3] = {0, 0, 1, };
OSQPInt P_p[3] = {0, 1, 3, };
OSQPFloat q[2] = {1.0, 1.0, };
OSQPFloat q_new[2] = {2.0, 3.0, };
OSQPFloat A_x[4] = {1.0, 1.0, 1.0, 1.0, };
OSQPInt A_nnz = 4;
OSQPInt A_i[4] = {0, 1, 0, 2, };
OSQPInt A_p[3] = {0, 2, 4, };
OSQPFloat l[3] = {1.0, 0.0, 0.0, };
OSQPFloat l_new[3] = {2.0, -1.0, -1.0, };
OSQPFloat u[3] = {1.0, 0.7, 0.7, };
OSQPFloat u_new[3] = {2.0, 2.5, 2.5, };
OSQPInt n = 2;
OSQPInt m = 3;
/* Exitflag */
OSQPInt exitflag = 0;
/* Solver, settings, matrices */
OSQPSolver *solver;
OSQPSettings *settings;
OSQPCscMatrix* P = malloc(sizeof(OSQPCscMatrix));
OSQPCscMatrix* A = malloc(sizeof(OSQPCscMatrix));
/* Populate matrices */
csc_set_data(A, m, n, A_nnz, A_x, A_i, A_p);
csc_set_data(P, n, n, P_nnz, P_x, P_i, P_p);
/* Set default settings */
settings = (OSQPSettings *)malloc(sizeof(OSQPSettings));
if (settings) osqp_set_default_settings(settings);
/* Setup solver */
exitflag = osqp_setup(&solver, P, q, A, l, u, m, n, settings);
/* Solve problem */
if (!exitflag) exitflag = osqp_solve(solver);
/* Update problem */
if (!exitflag) exitflag = osqp_update_data_vec(solver, q_new, l_new, u_new);
/* Solve updated problem */
if (!exitflag) exitflag = osqp_solve(work);
/* Cleanup */
osqp_cleanup(solver);
if (A) free(A);
if (P) free(P);
if (settings) free(settings);
return (int)exitflag;
};
③ Update matrices:更新矩阵P和A
#include <stdlib.h>
#include "osqp.h"
int main(int argc, char **argv) {
/* Load problem data */
OSQPFloat P_x[3] = {4.0, 1.0, 2.0, };
OSQPFloat P_x_new[3] = {5.0, 1.5, 1.0, };
OSQPInt P_nnz = 3;
OSQPInt P_i[3] = {0, 0, 1, };
OSQPInt P_p[3] = {0, 1, 3, };
OSQPFloat q[2] = {1.0, 1.0, };
OSQPFloat q_new[2] = {2.0, 3.0, };
OSQPFloat A_x[4] = {1.0, 1.0, 1.0, 1.0, };
OSQPFloat A_x_new[4] = {1.2, 1.5, 1.1, 0.8, };
OSQPInt A_nnz = 4;
OSQPInt A_i[4] = {0, 1, 0, 2, };
OSQPInt A_p[3] = {0, 2, 4, };
OSQPFloat l[3] = {1.0, 0.0, 0.0, };
OSQPFloat l_new[3] = {2.0, -1.0, -1.0, };
OSQPFloat u[3] = {1.0, 0.7, 0.7, };
OSQPFloat u_new[3] = {2.0, 2.5, 2.5, };
OSQPInt n = 2;
OSQPInt m = 3;
/* Exitflag */
OSQPInt exitflag = 0;
/* Solver, settings, matrices */
OSQPSolver *solver;
OSQPSettings *settings;
OSQPCscMatrix* P = malloc(sizeof(OSQPCscMatrix));
OSQPCscMatrix* A = malloc(sizeof(OSQPCscMatrix));
/* Populate matrices */
csc_set_data(A, m, n, A_nnz, A_x, A_i, A_p);
csc_set_data(P, n, n, P_nnz, P_x, P_i, P_p);
/* Set default settings */
settings = (OSQPSettings *)malloc(sizeof(OSQPSettings));
if (settings) osqp_set_default_settings(settings);
/* Setup solver */
exitflag = osqp_setup(&solver, P, q, A, l, u, m, n, settings);
/* Solve problem */
if (!exitflag) exitflag = osqp_solve(solver);
/* Update problem
NB: Update only upper triangular part of P
*/
if (!exitflag) exitflag = osqp_update_data_mat(solver,
P_x_new, OSQP_NULL, 3,
A_x_new, OSQP_NULL, 4);
/* Solve updated problem */
if (!exitflag) exitflag = osqp_solve(work);
/* Cleanup */
osqp_cleanup(solver);
if (A) free(A);
if (P) free(P);
if (settings) free(settings);
return (int)exitflag;
};
应用:
① Huber fitting
② Lasso
③ Least-squares:最小二乘法
④ Model predictive control (MPC)
我们考虑将线性时不变动力系统控制到某个参考状态的问题。 为了实现这一点,我们使用约束线性二次 MPC,它在每个时间步长求解以下有限视界最优控制问题
x
r
∈
R
n
x
x_r∈R^{n_x}
xr∈Rnx
⑤ Portfolio optimization
⑥ Support vector machine (SVM):支持向量机