search

LinaMiao / PQN-

1 stars 3 forks source link

task 11 : look careful into lasso problem #11

Open LinaMiao opened 12 years ago

LinaMiao commented 12 years ago

-- latest version of pqn -- for numerical example -- for seismic example

LinaMiao commented 12 years ago

-- well, minConf_PQN seems not converge very well, did I do something wrong to it? Have to go back and check.

-- will lbfgs converge without wolf line search?

LinaMiao commented 12 years ago

-- testing whether the original pqn can solve my lasso correctly

-- it turns out no, I'm not ruining things. The best performance of minConf_pqn is decreasing norm(r,2) all until to 1e-3. Then, directional derivative just less than opTol, if decreasing this opTol, or equally do not break when it happens, the function will still decrease the objective, just very slow.

-- And, making opTol in spg very small does not solve this problem, no matter how likely the direction is a Newton's one, still can not make things better

-- however, though norm of r is still not so small, f is around 1e-5. -- So maybe direction is not quite right

LinaMiao commented 12 years ago

-- by setting the opTol to be very small, like 1e-12(1e-6^2), working fine

LinaMiao commented 12 years ago

-- how about for seismic examples