search

IMTEK-Simulation / NuMPI

Utilities for MPI-parallel numerical calculations with Python
MIT License
2 stars 1 forks source link

LBFGS needs more iterations than scipy on 3DOF quadratic problem #31

Open sannant opened 4 years ago

sannant commented 4 years ago

Can be observed looking at the debug prints in test_3D

7caa863b

sannant commented 4 years ago

I tested this more carefully in the branch bug/20_high_number_iterations, in the files helpers/debug_compare_lbfgs_implementations_evaluations.py, and debug_compare_two_LBFGS_implementations.py

I think I was confused reading the number of iterations. The reason the MPI implementation suddenly needed more iterations is that there was no gtol prescribed in the test.

The mpi LBFGS is actually slightly faster than the old one. grafik

However scipy is still twice as fast on this particular problem. So there is still something suboptimal happening.

grafik the three plots show the evolution of the 3 degrees of freedom.

Differences can be expected since we have not implemented the same line-search algorithm than on scipy.