Bug fix in parallel processing for main driver and
mlmc.test (thanks to Qian Xin, University of Bristol, for
bug report).
At the same time, improve the method of splitting simulations in
parallel for the main mlmc driver, so that work is more
evenly distributed to keep all cores busy.
mlmc 2.1.0
Add parameter value checks in mlmc.test.
Allow user to specify alpha, beta, and
gamma to mlmc.test, rather than forcing
estimation by linear regression. Note this is a departure from the
original Matlab code, but if they are left unspecified then the same
results as under Matlab are reproduced.
Improve specificity of some argument documentation in
mlmc.test.
mlmc 2.0.2
Package was removed from CRAN because I didn’t notice my old Oxford
email address wasn’t forwarding any longer. In order to comply with CRAN
changes, the C++ routines are now registered and maintainer info updated
to my Durham email.
The Matlab driver code by Mike Giles has been quite substantially
updated, so this major version bump in the R package addresses updating
this code to match the new driver API. None of these sub-bullets are bug
fixes, merely changing to match the new best-practice for the MLMC
driver designed by Mike Giles. In particular:
User level sampling functions must now also return the total cost of
all samples simulated at that level. Therefore user level sampler
functions must return a list with a sums and
cost element.
The gamma argument is no longer required, since it is
not used in automatic cost computation, and can be estimated as for
alpha and beta.
mlmc.test() no longer takes M, a level
refinement factor, since this was only used to calculate the cost as
N*M^l. Per above comment, the user now defines cost
completely via the return from the level sampler function.
Along these lines, mlmc.test() now uses the user
returned cost in all places: previously CPU time was measured as cost in
the convergence tests section, whilst the MLMC complexity tests
previously forced costs to be N*M^l.
Some (very) old bugs were squashed in the Euler-Maruyama
discretisation level sampler, opre_l() which affected
lookback call and Heston model options.
I managed to get hold of a Matlab license, so have now confirmed
that the examples in the docs return (within sampling variability) the
same results for both Euler-Maruyama and Milstein discretisation example
level sampler functions.
There is now a hex sticker! It is hopefully fairly self explanatory:
many fast simulations are done at low levels (lots of dice, with the
hare running at the bottom of the stairs); fewer simulations are done at
higher levels (fewer dice as you go up each step, with a tortoise and
fewest dice on top step)!