Made the solver use an SMO type iteration in the beginning before

switching to projected gradient steps.
pull/18/head
Davis King 9 years ago
parent 1a494a912b
commit c290381bf0

@ -118,6 +118,7 @@ namespace dlib
for (unsigned long c = 0; c < horizon; ++c)
{
lambda += trace(trans(B)*temp*B);
Q_diag[horizon-c-1] = diag(trans(B)*temp*B);
temp = trans(A)*temp*A + diagm(Q);
}
@ -264,9 +265,6 @@ namespace dlib
for (unsigned long i = 0; i < horizon; ++i)
MM[i] = trans(B)*M[i];
for (unsigned long i = 0; i < horizon; ++i)
v[i] = controls[i];
unsigned long iter = 0;
@ -289,6 +287,8 @@ namespace dlib
// Check the stopping condition, which is the magnitude of the largest element
// of the gradient.
double max_df = 0;
unsigned long max_t = 0;
long max_v = 0;
for (unsigned long i = 0; i < horizon; ++i)
{
for (long j = 0; j < controls[i].size(); ++j)
@ -298,7 +298,12 @@ namespace dlib
if (!((controls[i](j) <= lower(j) && df[i](j) > 0) ||
(controls[i](j) >= upper(j) && df[i](j) < 0)))
{
max_df = std::max(max_df, std::abs(df[i](j)));
if (std::abs(df[i](j)) > max_df)
{
max_df = std::abs(df[i](j));
max_t = i;
max_v = j;
}
}
}
}
@ -306,12 +311,37 @@ namespace dlib
break;
// take a step based on the gradient
for (unsigned long i = 0; i < horizon; ++i)
// We will start out by doing a little bit of coordinate descent because it
// allows us to optimize individual variables exactly. Since we are warm
// starting each iteration with a really good solution this helps speed
// things up a lot.
const unsigned long smo_iters = 50;
if (iter < smo_iters)
{
if (Q_diag[max_t](max_v) == 0) continue;
// Take the optimal step but just for one variable.
controls[max_t](max_v) = -(df[max_t](max_v)-Q_diag[max_t](max_v)*controls[max_t](max_v))/Q_diag[max_t](max_v);
controls[max_t](max_v) = put_in_range(lower(max_v), upper(max_v), controls[max_t](max_v));
// If this is the last SMO iteration then don't forget to initialize v
// for the gradient steps.
if (iter+1 == smo_iters)
{
for (unsigned long i = 0; i < horizon; ++i)
v[i] = controls[i];
}
}
else
{
v_old[i] = v[i];
v[i] = clamp(controls[i] - 1.0/lambda * df[i], lower, upper);
controls[i] = clamp(v[i] + (std::sqrt(lambda)-1)/(std::sqrt(lambda)+1)*(v[i]-v_old[i]), lower, upper);
// Take a projected gradient step.
for (unsigned long i = 0; i < horizon; ++i)
{
v_old[i] = v[i];
v[i] = clamp(controls[i] - 1.0/lambda * df[i], lower, upper);
controls[i] = clamp(v[i] + (std::sqrt(lambda)-1)/(std::sqrt(lambda)+1)*(v[i]-v_old[i]), lower, upper);
}
}
}
}
@ -329,6 +359,7 @@ namespace dlib
matrix<double,S,1> target[horizon];
double lambda; // abound on the largest eigenvalue of the hessian matrix.
matrix<double,I,1> Q_diag[horizon];
matrix<double,I,1> controls[horizon];
};

Loading…
Cancel
Save