// Copyright (C) 2015 Davis E. King (davis@dlib.net)
// License: Boost Software License See LICENSE.txt for the full license.
#include <string>
#include <sstream>
#include <dlib/control.h>
#include <dlib/optimization.h>
#include "tester.h"
namespace
{
using namespace test;
using namespace std;
using namespace dlib;
logger dlog("test.mpc");
template <
typename EXP1,
typename EXP2,
typename T, long NR, long NC, typename MM, typename L
>
unsigned long solve_qp_box_using_smo (
const matrix_exp<EXP1>& _Q,
const matrix_exp<EXP2>& _b,
matrix<T,NR,NC,MM,L>& alpha,
matrix<T,NR,NC,MM,L>& lower,
matrix<T,NR,NC,MM,L>& upper,
T eps,
unsigned long max_iter
)
/*!
ensures
- solves: 0.5*trans(x)*Q*x + trans(b)*x where x is box constrained.
!*/
{
const_temp_matrix<EXP1> Q(_Q);
const_temp_matrix<EXP2> b(_b);
//cout << "IN QP SOLVER" << endl;
//cout << "max eig: " << max(real_eigenvalues(Q)) << endl;
//cout << "min eig: " << min(real_eigenvalues(Q)) << endl;
//cout << "Q: \n" << Q << endl;
//cout << "b: \n" << b << endl;
// make sure requires clause is not broken
DLIB_ASSERT(Q.nr() == Q.nc() &&
alpha.size() == lower.size() &&
alpha.size() == upper.size() &&
is_col_vector(b) &&
is_col_vector(alpha) &&
is_col_vector(lower) &&
is_col_vector(upper) &&
b.size() == alpha.size() &&
b.size() == Q.nr() &&
alpha.size() > 0 &&
0 <= min(alpha-lower) &&
0 <= max(upper-alpha) &&
eps > 0 &&
max_iter > 0,
"\t unsigned long solve_qp_box_using_smo()"
<< "\n\t Invalid arguments were given to this function"
<< "\n\t Q.nr(): " << Q.nr()
<< "\n\t Q.nc(): " << Q.nc()
<< "\n\t is_col_vector(b): " << is_col_vector(b)
<< "\n\t is_col_vector(alpha): " << is_col_vector(alpha)
<< "\n\t is_col_vector(lower): " << is_col_vector(lower)
<< "\n\t is_col_vector(upper): " << is_col_vector(upper)
<< "\n\t b.size(): " << b.size()
<< "\n\t alpha.size(): " << alpha.size()
<< "\n\t lower.size(): " << lower.size()
<< "\n\t upper.size(): " << upper.size()
<< "\n\t Q.nr(): " << Q.nr()
<< "\n\t min(alpha-lower): " << min(alpha-lower)
<< "\n\t max(upper-alpha): " << max(upper-alpha)
<< "\n\t eps: " << eps
<< "\n\t max_iter: " << max_iter
);
// Compute f'(alpha) (i.e. the gradient of f(alpha)) for the current alpha.
matrix<T,NR,NC,MM,L> df = Q*alpha + b;
matrix<T,NR,NC,MM,L> QQ = reciprocal_max(diag(Q));
unsigned long iter = 0;
for (; iter < max_iter; ++iter)
{
T max_df = 0;
long best_r =0;
for (long r = 0; r < Q.nr(); ++r)
{
if (alpha(r) <= lower(r) && df(r) > 0)
;//alpha(r) = lower(r);
else if (alpha(r) >= upper(r) && df(r) < 0)
;//alpha(r) = upper(r);
else if (std::abs(df(r)) > max_df)
{
best_r = r;
max_df = std::abs(df(r));
}
}
//for (long r = 0; r < Q.nr(); ++r)
long r = best_r;
{
const T old_alpha = alpha(r);
alpha(r) = -(df(r)-Q(r,r)*alpha(r))*QQ(r);
if (alpha(r) < lower(r))
alpha(r) = lower(r);
else if (alpha(r) > upper(r))
alpha(r) = upper(r);
const T delta = old_alpha-alpha(r);
// Now update the gradient. We will perform the equivalent of: df = Q*alpha + b;
for(long k = 0; k < df.nr(); ++k)
df(k) -= Q(r,k)*delta;
}
if (max_df < eps)
break;
}
//cout << "df: \n" << trans(df) << endl;
//cout << "objective value: " << 0.5*trans(alpha)*Q*alpha + trans(b)*alpha << endl;
return iter+1;
}
// ----------------------------------------------------------------------------------------
namespace impl_mpc
{
template <long N>
void pack(
matrix<double,0,1>& out,
const std::vector<matrix<double,N,1> >& item
)
{
DLIB_CASSERT(item.size() != 0,"");
out.set_size(item.size()*item[0].size());
long j = 0;
for (unsigned long i = 0; i < item.size(); ++i)
for (long r = 0; r < item[i].size(); ++r)
out(j++) = item[i](r);
}
template <long N>
void pack(
matrix<double,0,1>& out,
const matrix<double,N,1>& item,
const long num
)
{
out.set_size(item.size()*num);
long j = 0;
for (long r = 0; r < num; ++r)
for (long i = 0; i < item.size(); ++i)
out(j++) = item(i);
}
template <long N>
void unpack(
std::vector<matrix<double,N,1> >& out,
const matrix<double,0,1>& item
)
{
DLIB_CASSERT(out.size() != 0,"");
DLIB_CASSERT((long)out.size()*out[0].size() == item.size(),"");
long j = 0;
for (unsigned long i = 0; i < out.size(); ++i)
for (long r = 0; r < out[i].size(); ++r)
out[i](r) = item(j++);
}
}
template <long S, long I>
unsigned long solve_linear_mpc (
const matrix<double,S,S>& A,
const matrix<double,S,I>& B,
const matrix<double,S,1>& C,
const matrix<double,S,1>& Q,
const matrix<double,I,1>& R,
const matrix<double,I,1>& _lower,
const matrix<double,I,1>& _upper,
const std::vector<matrix<double,S,1> >& target,
const matrix<double,S,1>& initial_state,
std::vector<matrix<double,I,1> >& controls // input and output
)
{
using namespace impl_mpc;
DLIB_CASSERT(target.size() == controls.size(),"");
matrix<double> K(B.nr()*controls.size(), B.nc()*controls.size());
matrix<double,0,1> M(B.nr()*controls.size());
// compute powers of A: Apow[i] == A^i
std::vector<matrix<double,S,S> > Apow(controls.size());
Apow[0] = identity_matrix(A);
for (unsigned long i = 1; i < Apow.size(); ++i)
Apow[i] = A*Apow[i-1];
// fill in K
K = 0;
for (unsigned long r = 0; r < controls.size(); ++r)
for (unsigned long c = 0; c <= r; ++c)
set_subm(K,r*B.nr(),c*B.nc(), B.nr(), B.nc()) = Apow[r-c]*B;
// fill in M
set_subm(M,0*A.nr(),0,A.nr(),1) = A*initial_state + C;
for (unsigned long i = 1; i < controls.size(); ++i)
set_subm(M,i*A.nr(),0,A.nr(),1) = A*subm(M,(i-1)*A.nr(),0,A.nr(),1) + C;
//cout << "M: \n" << M << endl;
//cout << "K: \n" << K << endl;
matrix<double,0,1> t, v, lower, upper;
pack(t, target);
pack(v, controls);
pack(lower, _lower, controls.size());
pack(upper, _upper, controls.size());
matrix<double> QQ(K.nr(),K.nr()), RR(K.nc(),K.nc());
QQ = 0;
RR = 0;
for (unsigned long c = 0; c < controls.size(); ++c)
{
set_subm(QQ,c*Q.nr(),c*Q.nr(),Q.nr(),Q.nr()) = diagm(Q);
set_subm(RR,c*R.nr(),c*R.nr(),R.nr(),R.nr()) = diagm(R);
}
matrix<double> m1 = trans(K)*QQ*K+RR;
matrix<double> m2 = trans(K)*QQ*(M-t);
// run the solver...
unsigned long iter;
iter = solve_qp_box_using_smo(
m1,
m2,
v,
lower,
upper,
0.00000001,
100000);
//cout << "iterations: " << iter << endl;
unpack(controls, v);
return iter;
}
void test_with_positive_target_error_thresh()
{
// a basic position + velocity model
matrix<double,2,2> A;
A = 1, 1,
0, 1;
matrix<double,2,1> B, C;
B = 0,
1;
C = 0.0,0.0; // no constant bias
matrix<double,2,1> Q;
Q = 2, 0; // only care about getting the position right
matrix<double,1,1> R, lower, upper;
R = 0.001;
// We set it up so that the controller that only cares about getting to the target but
// doesn't care about what happens after can just give large controls of -1. It will fly
// past the target once it gets there, but it doesn't care about that. The solver_normal
// however wants to come to rest at the target. So it will have to give a much smaller
// control to get it moving towards the target and then slowly decelerate over time to end
// up stopping. So it will take longer to get to the target, which we test below.
lower = -1.0;
upper = 0.05;
mpc<2,1,30> solver_normal(A,B,C,Q,R,lower,upper);
solver_normal.set_epsilon(0.00000001);
solver_normal.set_max_iterations(10000);
mpc<2,1,30> solver_target_thresh = solver_normal;
solver_target_thresh.set_target_error_threshold(1.0);
matrix<double,2,1> state_normal;
state_normal = 10, 0;
matrix<double,2,1> state_target_thresh = state_normal;
// Run the control law with and without set_target_error_threshold() called. We should
// observe that the controller using set_target_error_threshold(1) drives us to the target
// state faster, however, only the normal solver results in a rest state at the target
// state.
int time_at_target_normal = -1;
int time_at_target_other = -1;
for (int i = 0; i < 30; ++i)
{
print_spinner();
const double normal_control = solver_normal(state_normal);
state_normal = A*state_normal + B*normal_control + C;
const double target_thresh_control = solver_target_thresh(state_target_thresh);
state_target_thresh = A*state_target_thresh + B*target_thresh_control + C;
const double normal_error = trans(state_normal)*diagm(Q)*state_normal;
const double target_error = trans(state_target_thresh)*diagm(Q)*state_target_thresh;
if (normal_error < 1.0 && time_at_target_normal == -1)
time_at_target_normal = i;
if (target_error < 1.0 && time_at_target_other == -1)
time_at_target_other = i;
}
// Check that the normal solver took longer to get to the target.
DLIB_TEST(time_at_target_normal == 13);
DLIB_TEST(time_at_target_other == 7);
// Check that the normal solver ends up right on the target.
DLIB_TEST(length(state_normal) < 0.01);
// But the one that got there faster blew past the target and is no way off in the distance.
DLIB_TEST(length(state_target_thresh) > 20);
}
class test_mpc : public tester
{
public:
test_mpc (
) :
tester ("test_mpc",
"Runs tests on the mpc object.")
{}
void perform_test (
)
{
// a basic position + velocity model
matrix<double,2,2> A;
A = 1, 1,
0, 1;
matrix<double,2,1> B, C;
B = 0,
1;
C = 0.02,0.1; // no constant bias
matrix<double,2,1> Q;
Q = 2, 0; // only care about getting the position right
matrix<double,1,1> R, lower, upper;
R = 1;
lower = -0.2;
upper = 0.2;
std::vector<matrix<double,1,1> > controls(30);
std::vector<matrix<double,2,1> > target(30);
for (unsigned long i = 0; i < controls.size(); ++i)
{
controls[i] = 0;
target[i] = 0;
}
mpc<2,1,30> solver(A,B,C,Q,R,lower,upper);
solver.set_epsilon(0.00000001);
solver.set_max_iterations(10000);
matrix<double,2,1> initial_state;
initial_state = 0;
initial_state(0) = 5;
for (int i = 0; i < 30; ++i)
{
print_spinner();
matrix<double,1,1> control = solver(initial_state);
for (unsigned long i = 1; i < controls.size(); ++i)
controls[i-1] = controls[i];
// Compute the correct control via SMO and make sure it matches.
solve_linear_mpc(A,B,C,Q,R,lower,upper, target, initial_state, controls);
dlog << LINFO << "ERROR: " << length(control-controls[0]);
DLIB_TEST(length(control-controls[0]) < 1e-7);
initial_state = A*initial_state + B*control + C;
//cout << control(0) << "\t" << trans(initial_state);
}
{
// also just generally test our QP solver.
matrix<double,20,20> Q = gaussian_randm(20,20,5);
Q = Q*trans(Q);
matrix<double,20,1> b = randm(20,1)-0.5;
matrix<double,20,1> alpha, lower, upper, alpha2;
alpha = 0;
alpha2 = 0;
lower = -4;
upper = 3;
solve_qp_box_using_smo(Q,b,alpha,lower, upper, 1e-12, 500000);
solve_qp_box_constrained(Q,b,alpha2,lower, upper, 1e-12, 50000);
dlog << LINFO << trans(alpha);
dlog << LINFO << trans(alpha2);
dlog << LINFO << "objective value: " << 0.5*trans(alpha)*Q*alpha + trans(b)*alpha;
dlog << LINFO << "objective value2: " << 0.5*trans(alpha2)*Q*alpha + trans(b)*alpha2;
DLIB_TEST_MSG(max(abs(alpha-alpha2)) < 1e-7, max(abs(alpha-alpha2)));
}
test_with_positive_target_error_thresh();
}
} a;
}