hasty-hamiltonian

Speedy gradient-based traversal through parameter space.
git clone git://git.jtobin.io/hasty-hamiltonian.git
Log | Files | Refs | README | LICENSE

README.md (1101B)

      1 # hasty-hamiltonian
      2 
      3 Speedy, gradient-based traversal through parameter space.
      4 
      5 Exports a `mcmc` function that prints a trace to stdout, a `chain` function for
      6 collecting results in memory, and a `hamiltonian` transition operator that can
      7 be used more generally.
      8 
      9 If you don't want to calculate your gradients by hand you can use the handy
     10 [ad](https://hackage.haskell.org/package/ad) library for automatic
     11 differentiation.
     12 
     13     import Numeric.AD (grad)
     14     import Numeric.MCMC.Hamiltonian
     15 
     16     target :: RealFloat a => [a] -> a
     17     target [x0, x1] = negate ((x0 + 2 * x1 - 7) ^ 2 + (2 * x0 + x1 - 5) ^ 2)
     18 
     19     gTarget :: [Double] -> [Double]
     20     gTarget = grad target
     21 
     22     booth :: Target [Double]
     23     booth = Target target (Just gTarget)
     24 
     25     main :: IO ()
     26     main = withSystemRandom . asGenIO $ mcmc 10000 0.05 20 [0, 0] booth
     27 
     28 ![trace](https://dl.dropboxusercontent.com/spa/u0s6617yxinm2ca/h6ty39dl.png)
     29 
     30 *hasty-hamiltonian* is a member of the [declarative][decl] suite of libraries,
     31 containing a bunch of MCMC algorithms that play nicely together.
     32 
     33 [decl]: https://github.com/jtobin/declarative