Tuesday -- two insights

Control & Optimization

Dominique Bonvin visit

  • two-step control (optimize parameters, optimize mean-square error between model and measurement also by parameter adjustment) fails when reality isn’t in the model set.  Problem is under-determined (N free variables, 4N equations).  (converges, but onto the wrong optimum).

  • Better solution – adjust the penalty/regularization terms to to get better agreement between model and data

  • Bonvin approach – adjust the control instead (NCO Control).  Can be almost model free in the control process.   (**)

pdg-control Training Problem 1b & 2

  • Coded SDP with costs to changing quotas.

  • Best way to visualize the optimal solution?

Evolution

  • Hessian calculation?  Non-trivial – how to get Hessian from a non-gradient method, or a gradient method when you need to handle bounds?  Various numerical challenges.

  • Wrightscape model comparisons – comparing meaningful hypotheses! (**)

Day’s graphs log

[flickr-gallery mode=“search” tags=“phylogenetics, PDG_Control” min_upload_date=“2011-12-06 7:00:37” max_upload_date=“2011-12-06 23:23:37”]