Notes on AD Model Builder by a novice
Contents
1. elementwise vector multiplication this is mostly a problem for R S PLUS users since C C has no such expectation and MATLAB uses for elementwise operations http courses washington edu fish507 ADMB 20living 20doc doc that describes how to turn on bounds checking but it looks like this only works in an IDE is there an equivalent of parallel min max pmin pmax in R It would be useful for bounds checking although the built in posfun does work various experimentation with bounds etc bounding parameters at zero where sensible seems to help even when it doesn t appear that there is such a problem at the end point tpl2rem needed for random effects models appears to be fussy about dos vs unix mode not sure about this any more don t know if I can replicate it One can apparently ignore errors of the form cat xxalloc4 tmp No such file or directory cat xxalloc5 tmp No such file or directory Error executing command cat xxglobal tmp xxhtop tmp header tmp xxalloci tmp xxalloc2 tmp xxalloc3 tmp xxalloc4 tmp xxallocd tmp xxtopm tmp xxalloc6 tmp gt mccoypred3 cpp I think these are produced by ADMB RE models that are too simple some warnings when compiling code many fewer as of a recent update would be nice if legal code compiled cleanly without warnings My ADMB models now compile cleanly but my ADMB RE code gives In file included from dfib2fun h 126 from m2. Notes on AD Model Builder by a novice Ben Bolker November 10 2010 Mostly created 11 August 2009 not significantly updated since then Some issues may have been resolved This document describes my discoveries and frustrations as I started to learn how to use AD Model Builder and ADMB RE Note to ADMB developers please take this all as constructive criticism ADMB is a great tool and I hope these comments will contribute to its development and ease of uses for others 1 I got quite confused by the distinction between tpl2cpp and tpl2rem I now know that these are used for constructing models without and with random effects respectively but this wasn t clear initially when I was poking around in the bin directory trying to figure out what to do There seem to be a confusing number of different possible scripts and wrappers the ADMB RE manual refers to an admb script which didn t come with the binary version of ADMB that I downloaded despite the fact that the binary did allow the construction of ADMB RE models I eventually downloaded from http admb project org community and got the admb script Appendix C of the ADMB RE documentation is really useful it might also be useful to mention adlink and adcomp which are used in the admb script a Makefile would be nice so one could rebuild only the necessary pieces but it might be overkill and might be too hard to make it properly cross platform since Windows users etc w
3. ccoypred5 cpp 2 df 32fun h 98 warning extra qualification df3_two_vector on member allocate df 32fun h 150 warning extra qualification df3_two_matrix on member df3_two_matrix In file included from mccoypred5 cpp 2 dfib2fun h 715 warning extra qualification df1ib2_gradlist on member write_pass1 12 13 In file included from df1b2fun h 1555 from mccoypred5 cpp 2 df3fun h 87 warning extra qualification df3_one_vector on member allocate df3fun h 139 warning extra qualification df3_one_matrix on member df3_one_matrix I had trouble deciding whether random effect values needed to be included as a starting value or not since they are specified as random_effects_vector rather than init_vector it wasn t clear to me Dave Fournier points out that they are not required my best guess is that the default behavior is to make all starting values equal to zero if a PIN file is not specified and if the values are not specified as bounded in the PARAMETER_SECTION and if no explicit assignments are made in the INITIALIZATION_SECTION but it s a little hard to tell He says the easiest way to make a pin file is to run the model with the command line option maxfn 0 sets maximum num ber of function evaluations to zero and copy the output par file p01 in this case to pin and edit it P 1 28 of the ADMB manual says If the model finds the parameter file whose default name is admodel par Should
4. e authors haven t used at least S3 classes to define a class of admb or admb_fit objects and associated methods plot print summary etc maybe this is because of the GUI approach In any case I wouldn t want to reinvent the wheel too thoroughly 8 would be nice to have negative log likelihood helper functions defined for some other standard distributions Poisson Cauchy neg binomial are defined but not binomial beta binomial Beta 9 It might be useful to write some vignettes showing how one could use model matrix and perhaps some digging in the guts of lmer models to easily construct model matrices for the random and fixed effects Again wouldn t want to re invent too much of ghmmADMB 10 glmmADMB is nice but frustrated me a bit a documentation holes and general missing pieces e g some of the documentation still has stubs in it see help g1lmmADMB package offsets are implemented but not documented at least in version 0 3 I get a warning about cannot remove file nbmm std when I run example glmm admb b I find the banner Welcome to ghmmADMB irritating I prefer packages that don t announce themselves c the license is undefined d general binomial with N gt 1 variables aren t possible although the package description only says Poisson or negative binomial response distributions so even Bernoulli is better than documented e it prints all the in
5. mber of clicks installer for Win dows Does the IDE version put everything in the right places so that ADMB just works Can I get admb mode from the google code pages without getting the full 66MB zip file and digging through it how about deb rpm packages for Linux Debian repositories so I can just say apt get install admb it would be nice to have df1b2vector pow const dfib2vector amp v const dfib2variable amp x defined in ADMB RE so I could leave it out of my TPL file it seems a little odd that sdreport must be explicitly turned on to get MCMC results for a parameter when otherwise all parameters are re ported on is this just to save space when the parameter vector is long 7 Rintegration there are various attempts but I think IM perhaps NSHO that they could be improved and made more R ish I get the sense that there is not a huge amount of overlap alas between people who know ADMB and R idioms well The primary integration attempts are e some code in glmmADMB which I have made use of in my stuff espe cially dat_write and pin_write See below for nit picks e ADMB2R which e PBSadmb http code google com p pbs software this might be great but it s GUI and not cross platform both of which put me off a little bit I am used to developing in R and would like to work that way Glancing over the PBSadmb user s manual I m a little bit surprised that th
6. ouble const char 14 15 16 mccoypred5 cpp In member function virtual void df1b2_parameters allocate mccoypred5 cpp 238 error no matching function for call to df1b2_init_bounded_number allocate const char 2 dfib2fun h 1315 note candidates are void dfib2_init_bounded_number allocate double double int const char dfib2fun h 1318 note void dfib2_init_bounded_number allocate double double const char If I don t add likeprof number declarations running the code with the lprof switch doesn t do anything bad but doesn t seem to produce profiles either at least I can t find it check likelihood on scaled N 0 1 variables p 18 of ADMB RE man ual try this both ways non significant typos in manuals where should I report these note negative binomial log likelihood log negbinomial_density is pa rameterized with variance mean which is 1 u k or 1 0 in the more usual overdispersion parameter parameterizations Thoughts wishes 1 it will be nice to have a full SVN setup comparable to more typical nix open source situations where checking out the source directories and running configure make in the root directory Just Works At the moment I can get all of the source files but it s fairly difficult to figure out the appropriate flags etc etc for building and linking the appropriate library files any chance of a 1 click or small nu
7. ould have to get a make executable from somewhere Probably most of this confusion would have been avoided if the binary distribution I downloaded had actually had an admb script and if that was consistently referred to throughout both the ADMB and the ADMB RE manual ADMB manual only mentions tpl2cpp of course since it preceded ADMB RE ADMB RE mentions both but never explicitly says exactly what they are 2 At one point I thought that file names with dashes were somehow broken but now I think I must just have been confused can t replicate 3 ADMB is case sensitive in a slightly nasty way which must come from its Windows DOS heritage but is a little odd on a nix system In particular editing tools admb ide scripts linux zip view searchterm Scripts it converts file names to lower case if the TPL file is called Simple tpl it will look for a DAT file called simple dat and fail to find one called Simple dat If parameters are specified the order in the PIN file must match the order in which they are defined in the TPL file this is perhaps obvious but screwed me up a few times I have gotten lazy and used to the expectation that things are named and matched up according to their names Seems one could write an automatic checker for this reasonably easily ADMB does the right thing with scalar vector arithmetic and vec tor matrix and vector vector but one must remember to use elem_prod for
8. termediate sampling results I would rather have it be silent or at worst produce some kind of estimated progress bar although that could be a little tricky for a standard progress bar since we don t really know how long it s going to take but we could just print progress dots every few steps f the one time I tried to use importance sampling it failed granted I didn t try too hard I will mention that Dave Fournier immediately sent back ADMB code that would do what I wanted but since I wasn t ready to plunge into full ADMB at the time I didn t try it out There are plenty of other nice extensions e g allow MCMC retrieval of MCMC results that would be straightforward What are all the files that admb spits out admodel dep admodel hes binary hesscheck hessian bin eva log luu rhes bar par parameter input p01 parameter output b01 binary parameter output
9. that be admodel pin Likelihood profiles are likelihood profiles 1prof likeprof_number sup posed to work for ADMB RE models Is likeprof_number supposed to be used in addition to the standard init_ definition or in its place Where is the likelihood profile information stored If I add a likeprof_number declaration to my non RE code I get In file included from mccoypred2 cpp 6 mccoypred2 htp 52 error declaration of param_likeprof_number model_parameters g mccoypred2 htp 49 error conflicts with previous declaration param_init_number model_parameters g With RE code I get In file included from mccoypred5 cpp 24 mccoypred5 htp 72 error declaration of param_likeprof_number model_parameters g mccoypred5 htp 59 error conflicts with previous declaration param_init_bounded_number model_parameters g mccoypred5 htp 126 error declaration of dfib2variable df1b2_parameters g mccoypred5 htp 113 error conflicts with previous declaration df1b2_init_bounded_number df1ib2_parameters g mccoypred5 cpp In constructor model_parameters model_parameters int int char mccoypred5 cpp 68 error no matching function for call to param_init_bounded_number allocate const char 2 usr local src admb include admodel h 1012 note candidates are void param_init_bounded_number allocate double double int const char usr local src admb include admodel h 1013 note void param_init_bounded_number allocate double d
Download Pdf Manuals
Related Search