User:Carl Boettiger/Notebook/Comparative Phylogenetics/2010/06/04
From OpenWetWare
Comparative Phylogenetics  Main project page Previous entry Next entry 
Comparative Phylogenetics TalkPractice talk in Wainwright lab meeting next Weds. 12 minutes.
Frameworks for approaching regimes question
Returning to Wrightscape approachHaven't worked on the full problem of inferring multiple peaks directly from comparative data using the joint probability calculation since the beginning of this notebook, so about time to return to the direct problem and compare with the more heuristic approach of "phylogenetically corrected clustering" and also synthesize the work on uncertainty which has been a dominate theme thus far (add category label for this group!) The direct calculation is still going by the name wrightscape, and the project (without its svn history unforunately) has been migrated to git and pushed to github. So thus far the current code for this project is independent of that in ComparativePhylogenetics project repository. This is the home of the Wrightscape repository. Since it's been a while since I've worked on this project, I begin by refreshing myself as to the organizational structure of the code. The callgraph provides a good summary: Generated with valgrind and visualized in kcachegrind by: make debug valgrind tool=callgrind ./jointprob kcachegrind callgrind.out.* kcachegrind then displays the call graph, (and other information), which can save as png.
analytic bm llik = 8.648713 analytic ou llik = 8.570838 matrix llik = 8.582334 analytic bm llik = 65.106558 analytic ou llik = 61.906121 matrix llik = 61.927070 So what is the matrix approach implement exactly? Desending the call graph further, we see it is mostly computing the transition matrix that sits at each node (which it can do directly using the analytic OU solution). Then the matrix solution to the joint probability is specified by: So what generality is gained by this approach? While the matrix formulation is general, this requires the direct calculation applied to each node. Should be able to simultaneously solve everything in a particular regime directly, if there is a meaningful way to MCMC over larger regime partitions rather than over node assignments. Evolution meeting
