Methods to optimize sample configurations using spatial simulated annealing. Multiple objective functions are implemented for various purposes, such as variogram estimation, spatial trend estimation and spatial interpolation. A general purpose spatial simulated annealing function enables the user to define his/her own objective function. Solutions for augmenting existing sample configurations and solving multi-objective optimization problems are available as well.

Optimization of Sample Configurations using Spatial Simulated Annealing

To get the current released version from CRAN:

`install.packages("spsann", dependencies = TRUE)`

To get the current development version from GitHub (you may have to install
the `devtools`

package):

`# install.packages("devtools")devtools::install_github("samuel-rosa/spsann")`

We use the *fork & pull* collaborative development model. This means that
anyone can fork the `spsann`

repository and push changes to their personal fork
without requiring access be granted to the source `spsann`

repository. The
changes will be reviewed and, if pertinent, will be pulled into the source
`spsann`

repository by the project maintainers.

We use the *fork & pull* collaborative development model mainly because it
reduces the amount of friction for new contributors. It makes it easier for
others to contribute. Besides, it is very popular with open source projects
because it allows people to work independently without upfront coordination.
Visit https://help.github.com/articles/using-pull-requests/ for more details on
how to fork a repository and make pull requests.

Now * spsann* can be used to augment an existing sample configuration, that is, add new sampling points
to a spatial sample configuration generated using

`optim...()`

functions, the user must pass to the function argument
`points`

an object of class `list`

containing two named sub-arguments: `fixed`

, a matrix with the
coordinates of the existing sample configuration -- kept fixed during the optimization --, and `free`

,
the number of sample points that should be added to the existing sample configuration -- free to move around
during the optimization.This is a major release of package * spsann* that includes several conceptual changes. Despite our efforts,
it was not possible to guarantee the compatibility with previous versions. We have decided not to deprecate
functions and function arguments because (1) this would require deprecating a lot of code and (2) you should
first read the updated package documentation to understand the conceptual changes that we have made before you
start using it. This is a summary of the changes:

- A completely new annealing schedule was implemented. The reason for this modification is that the former
annealing schedule showed to be inefficient during our tests. The new annealing schedule is the very simple
and most-used schedule proposed by Kirkpatrick et al. (1983). We have also replaced the acceptance criterion
with the well-known Metropolis criterion. This new implementation showed to be more efficient in our tests
than our early implementation. Setting up this new annealing schedule is done using the new function
`scheduleSPSANN`

. - A more elegant solution to jitter the sample points was implemented. It consists of using a finite set of candidate locations that are seen by the algorithm as the centre of grid cells. In the first stage, we select a grid cell with replacement. In the second stage, we select a location within that grid cell using simple random sampling. This guarantees that any location in the sampling region is a candidate location for the jittered sample point.
- Solving multi-objective combinatorial optimization problems (MOCOP) has become easier with the creation of
the new function
`minmaxPareto()`

. This function computes the Pareto maximum and minimum values of the objective functions that compose the MOCOP needed to scale the objective functions to the same approximate range of values. - The user can now chose to follow the progress of the optimization using a text progress bar in the R console
or a Tk progress bar widget. A Tk progress bar widget is useful when running
in parallel processors.**spsann** - The output of the optimization is now stored in an object of class
`OptimizedSampleConfiguration`

. This object contains three slots. The first (`points`

) holds the coordinates of the optimized sample configuration. The second,`spsann`

, stores information about the settings used with the spatial simulated annealing algorithm. The third,`objective`

, holds the settings used with the chosen objective function. Methods were implemented to retrieve information from the new class, as well as producing plots of the optimized sample configuration. - Package documentation was expanded and adapted to cope with the conceptual changes that were made. It also includes a vignette that gives a short description of the package and its structure, as well as presents a few examples on how to use the package. It is strongly recommended to read the new package documentation and the accompanying vignette before you start using the package.
- Finally, bugs were fixed, warning messages were improved, and a faster code was implemented whenever possible.

`devel`

branch was merged into`master`

branch.

- Package documentation was expanded. It now includes a vignette that gives a short description of the
package and its structure. The vignette also contains a few examples on how to use the package.
`knitr`

is the engine used to produce the package vignette.

- Documentation was expanded.

- Created S3 methods for extracting the objective function value and plotting an object of class
`OptimizedSampleConfiguration`

. - The class
`OptimizedSampleConfiguration`

is no longer exported. - Documentation was expanded.

- FIX: the computation of the number of point-pairs per lag-distance class in
`optimPPL`

was incorrect because it neglected the fact that, in a full distance matrix, two points*a*and*b*form two pairs, i.e.*ab*and*ba*. The mistake is due to the fact that we use`SpatialTools::dist1`

to compute the distance matrix instead of`stats::dist`

. - FEATURE: using a faster code to compute the number of points and point-pairs
per lag-distance class in
`optimPPL`

.

- Fixed minor bugs.

- Improved the warning message printed when converting numeric covariates into factor covariates.
- Created a new
`autofun`

to check the number of accepted jitters in the first chain. If the number of accepted jitters is superior to the value passed to`schedule$initial.acceptance`

, the process continues and a message is printed informing the proportion of jitters that have been accepted. - Included scaling factors in two of the objective functions of
`optimCLHS()`

following the original Fortran code of Budiman Minasny. - Use grey colours in plot with energy states; using only different line types was not enough to see the different lines -- using different colours makes it easier to see the differences among lines that represent different objective functions.
- Fixed minor bugs.

- The user can now chose the type of progress bar that should be used, with
options
`"txt"`

, for a text progress bar in the R console,`"tk"`

, to put up a Tk progress bar widget, and`NULL`

to omit the progress bar. A Tk progress bar widget is useful when runningin parallel processors. The**spsann**-package is now a suggested package.**tcltk** - Now we use grey colours to in the plot with the energy states.

- Solved NOTEs produced during CRAN check due to the use of functions from
default packages other than
`base`

, and due to examples that take more than 5 seconds to run.

- Created a function to plot the optimized sample configuration (
`plotOSC()`

), with options to display the evolution of the energy state and/or the optimized sample configuration. - The function used to compute the Pareto maximum and minimum values (
`minmaxPareto()`

) was optimized to be used with both ACDC and SPAN.

- Create a class (
`OptimizedSampleConfiguration`

) to store the output of`optim`

functions.

- The trick included in the
`optimMKV()`

-function to avoid errors due to the LDLfactor error of the-package had to be reformulated. We are now using**gstat**`try()`

with a default value which is returned in case of error.

- A completely new annealing schedule was implemented. The reason for this
modification is that the former annealing schedule, which was based on the
-package, showed to be inefficient during our tests. The new annealing schedule is the very simple and most-used schedule proposed by Kirkpatrick et al. (1983). We have also replaced the acceptance criterion used in the**intamapInteractive**-package with the well-known Metropolis criterion. This new implementation showed to be more efficient in our tests than our early implementation.**intamapInteractive** - Implementing a new annealing schedule and a new acceptance criterion required a moderate modification of the source code. Despite our efforts, it was not possible to guarantee the compatibility with previous versions.
- A new function was created to set up the annealing schedule:
`scheduleSPSANN()`

. - We are now using a more elegant solution to jitter the sample points. It
consists of using a finite set of candidate locations that are seen by the
algorithm as the centre of grid cells. In the first stage, we select a grid
cell with replacement. In the second stage, we select a location within that
grid cell using simple random sampling. This is the sampling method
implemented in the
-package.**spcosa** - The documentation of all functions has been fine tuned.
- A trick was included in the
`optimMKV()`

-function to avoid errors due to the LDLfactor error of the-package.**gstat** - There also is a new function to compute the Pareto maximum and minimum values of the objective functions
that compose a multi-objective optimization problem (MOOP):
`minmaxPareto()`

.

- Now
`x.max`

and`y.max`

are, by default, set to half of the maximum distance in the x- and y-coordinates of`candi`

, respectively. In the same manner, the argument`cutoff`

of`optimPPL()`

is set, by default, to half of the diagonal of the rectangle of sides`x.max`

and`y.max`

.

- Corrected a bug in
`optimCORR()`

that was causing the following error: Error in if (new_energy <= old_energy) { : missing value where TRUE/FALSE needed. This bug used to affect`optimACDC()`

and`optimSPAN()`

. - Created a version of the method proposed by Minasny and McBratney (2006),
known as the conditioned Latin hypercube sampling (
`optimCLHS`

). - Improved and updated the documentation of several functions.
- Improved the plotting functionality: each plot (the evolution of the energy state and the new system configuration) is now displayed in a separate device. This allows for a better visualization and allows the user to focus on a single plot if so s/he wishes.

- Improved and updated documentation.
is not a dependence any more.**gstat**- Fixed breaks due to changes in dependencies (
).**pedometrics**

- Submission to CRAN.

- An auxiliary function (
`objSPSANN()`

) was created to retrieve the energy state of an optimized sample configuration (OSC) at a given point of the optimization. - Long examples are not run any more to avoid overload of
`R CMD check`

. - The authors' list was updated with the respective roles.

- The documentation of all functions was significantly improved.
- Functions from default packages other than
are now imported to comply with the new change to the CRAN policy described at http://developer.r-project.org/blosxom.cgi/R-devel/NEWS/2015/06/29#n2015-06-29.**base**

- Using
`utils::globalVariables`

to avoid the`R CMD check`

note`no visible binding for global variable [variable name]`

. Source of the solution: http://stackoverflow.com/a/12429344/3365410. - The package
is not a dependency any more.**fields** - New default values were attributed to the following arguments:
`plotit`

,`track`

,`verbose`

, and`iteration`

. The first three were set to`FALSE`

, while the last was set to`100`

. `optimSPAN()`

and`objSPAN()`

are now full operational.

- Several internal function were renamed using a pattern that includes the name
of the respective objective function. For example,
`.optimPPLcheck()`

was renamed as`.checkPPL()`

, and`.getLagBreaks()`

was renamed as`.lagsPPL()`

. Note that the first part of the function name indicates what it does, while the second indicates the objective function to which it applies. This standardization is important to ease the construction of multi-objective optimization problems.

- Improvements in the family of ACDC, CORR, and DIST functions.
- Several pairs of internal function that were originally designed to deal with
different types of covariates (factor and numeric) were merged. Now a single
function does the job by using the key argument
`covars.type`

. - New internal functions now enable building multi-objective optimization problems more easily. They have also allowed to clean-up/simplify the source code.
- A new
`autofun`

was created to set-up the covariates (`covar`

).

- The
and**rgeos**packages are not dependencies any more.**plyr** - The
`boundary`

of the spatial domain can now be estimated internally. The user should use thepackage if a more precise**rgeos**`boundary`

is needed. - Now using a directory called 'R-autoFunction', where R code chunks that are
used in several functions of both families of
`obj...()`

and`optim...()`

functions are included in individual files. These R code chunks are used to automatically build internal functions. Currently, R code chunks are used to check the arguments of the family of`optim...()`

functions, prepare`points`

and`candi`

, set plotting options, estimate the`boundary`

, prepare for jittering, plot and jitter, and prepare the output. - BUG: the family of
`obj...()`

functions may not return the same criterion value of the optimized sample configuration returned by the family of`optim...()`

functions if the number of iterations used in the optimization is equal to 100. The problem seems to disappear if a larger number of iterations is used.

`spJitterFinite()`

now tries to find an alternative point if the new point already is included in the sample. The number of tries is equal to the total number of points included in the sample. Because the more points we have, the more likely it is that the candidate point already is included in the sample.

`spJitterFinite()`

now returns the old point if the new point already is in the sample. This is to avoid an infinite loop at the end of the optimization when the objective function creates a cluster of points.

- New version of
`optimACDC()`

, including new argument definitions; - In the multi-objective optimization problem case, now the graphical display includes the many objective functions being optimized along with the utility function.

- Special version designed for the course on Spatial Sampling for Mapping, 22 - 24 April 2015, Wageningen University Campus, Wageningen, The Netherlands, Under the auspices of the Graduate School for Production Ecology and Resource Conservation (PE&RC).

- new function to enable the user to define his/her own objective function;
- grammar check and enhanced documentation;

- new functions derived from
`optimACDC()`

:`optimDIST()`

and`optimCORR()`

; - new objective function: mean/maximum kriging variance;
- review of the family of PPL functions;
- using function tailored argument checking.

- in-development package;
- importing functions from the package
;**pedometrics** - preparing documentation.