Skip to content


Subversion checkout URL

You can clone with
Download ZIP
Branch: master
Fetching contributors…

Cannot retrieve contributors at this time

36 lines (34 sloc) 2.5 KB
2014 Mar 9:
- Huge gigantic change - updated so that the main branch enumerates structures and does MCMC on constants only!
- structure-learning is the old branch, modifying the programs. No shared code bewteen the two, to keep the constant branch very simple (this should change in the future)
2014 Mar 4:
- Changed to make mainly MCMC over constants
- removed MAPs so that a run either returns samples or MAPS, no separate MAP arrays
- Also no separate host_hypotheses, everything lives in mcmc_spec
- Added enumeration abilities, so you can enumerate all trees up to a given size, and run MCMC on the constants
2013 Nov 24:
- Fixed major bug in program insertion. Using slower code for now, but more clearly correct.
- Removed resampling (this worked with old interfaces)
- Added a repetition loop to allow multiple repeats within a run
2013 Nov 20:
- Major revision. Made kernel function take an "mcmc_specification" and return "mcmc_results", so that reading/writing to kernels is all bundled. Most kernels are not updated to handle this, so MH-simple-kernel is the only one currently working. The others may not be re-implemented.
- Support for constants has been broken (minorly) in this revision, and may or may not be added back in.
- Also make "x" much less likely in the prior, yielding much simpler expressions.
2013 Oct 26:
- Updates to postprocessing scripts, adding gaussian process and polynomial fits
2013 Sep 16:
- Fixed bug in ll scaling for annealing (sign flip)
2013 Sep 14:
- Split up multiple kernels into a separate directory
- Fixed readme
- fixed performance stats on memory copy
- added fancy temperature scheduling, default annealing
2013 Sep 10:
- Added make_primitives_header which allows different sets of primitives to be defined and built into the headers. This is a clumsy way to do it, but doing it with C macros seems to be a bit too much. This also involved moving computation of the prior (from expected length) into main.
2012 Sep 7:
- Fixed bug so that tops *and* samples are stored so that we don't over-mix chains (and all find the same top)
- Added a prior that penalizes deep x (variable) values
- Changed prior and likelihood to set the h->prior and h->likelihood values inside, rather than in posterior
2012 Sep 6:
- Fixed bug in sorting so that top really does not store duplicates (sorting by posterior is not enough since different expressions can have identical posteriors! This can give duplicates in the top list--as in PrimeCounting)
- Fixed bug in reporting performance stats
Jump to Line
Something went wrong with that request. Please try again.