NEWS.md

In andrewmarx/samc: Spatial Absorbing Markov Chains

samc 4.0.0

Breaking changes

• visitation_net(samc, origin, dest) has two changes to its behavior:
• It now behaves consistently with other metrics and returns a single value representing the result at the destination node. Previously, it behaved like the passage() function from gdistance and returned a vector using the origin and destination node information to perform a correction to the flow values.
• Previously, the function used the same mathematical implementation as passage() from gdistance. In retrospect, the final part of this implementation only works properly in the specific case of a single start point and a single end point, which is correct for the input requirements of passage(), but is not quite correct for when there are multiple nodes with absorption. The new version properly accounts for general cases involving multiple absorbing locations and varying intensities of absorption. This does change the results from older versions. The difference between the two versions is equivalent to one half the results from mortality(). In practice, landscapes with widespread absorption will see a negligible difference; in current large-scale examples, the effect is several magnitudes of order less than the net flow value. The difference may be more prominent when dealing with a very small number of absorbing nodes.

New features

• New input options for metrics
• visitation_net(samc, origin)
• visitation_net(samc, init)
• visitation_net(samc, init, dest)
• New precision option in the options list for samc(). Can choose between "single" and "double" precision
• Currently only applies to samc objects built for convolutions.
• Default is "double" and provides 15-16 digits of precision. This is the same as all previous versions.
• The "single" option provides 7-8 digits precision, but also reduces RAM use by roughly half. It's also slightly faster for metric calculations (~20% less time on one device), and also scales slightly better with multiple cores.

Other

• Enabled vector inputs for init parameters regardless of the data type used for inputs to the samc() function. This is intended to provide flexibility for special situations, and it's recommended that init inputs reflect the data types used for samc().
• Reorganization of internal functions.
• Improvements to model and options validation.

Bug fixes

• Fixed check() node count approach for convolution.
• Fixed convolution short-term metrics with multiple time inputs.

samc 3.3.0

New features

• New input options for metrics
• cond_passage(samc, init, dest)
• dispersal(samc, origin, dest, time)
• distribution(samc, init, dest, time)
• mortality(samc, init, dest, time)

Other

• Lot's of internal refactoring and consolidation.
• Various small memory and speed optimizations.
• Various small fixes.
• License update.

samc 3.2.1

Bug fix

• Fixed a minor issue in check() that was preventing use of the init parameter for samc objects created directly from a matrix.

samc 3.2.0

New Features

• Added the visitation_net() function for calculating net movement flow (vs total from visitation())

samc 3.1.0

New Features

• Added support for the convolution algorithm described by Hughes et al. (2023, DOI: 10.1007/s10980-023-01619-9):
• This algorithm is very memory efficient and fast, making it a good choice in memory-constrained scenarios.
• Not currently supported for all metrics.
• Only supported with metrics that use the init parameter.
• Long-term metrics have indeterminate run times that can be either really fast or really slow, depending on the data used to setup the model.
• Does not currently support NA values in landscape data.
• Only relevant for creating samc-class objects from raster/map data; does not work with manually constructed transition matrices.
• Added experimental support for correlated random-walk (CRW) models:
• Users should generally assume that the CRW walk requires an order of magnitude more memory than the default random-walk model.
• The current implementation relies on the von Mises distribution and allows users to specify a single global value for kappa as a turning probability. Future versions of the package will expand on this to support cell specific turning probabilities and directional bias.
• Currently only supports metrics with an origin input. It does not support the init or dest inputs. The origin input for CRW is a matrix with the cell number and a direction
• The map() function averages the results for all directions. This could change in the future to allow for more flexibility.

Performance

• Added support for built-in named transition functions, which results in significantly faster samc-class object creation by eliminating overhead associated with user-defined functions. Currently, only "1/mean(x)" is supported, but others can be added in the future.

Other

• Added the toy resistance data used for a workshop at the IALE 2021 conference.

samc 3.0.2

New Features

• Added support for the terra package for raster data. Internally, the package now uses terra and converts RasterLayer objects to SpatRaster objects. It's recommended that users switch to the terra package for loading and preparing raster data for samc.
• Made the rasterize() function publicly available. Mainly useful for converting matrices to a SpatRaster that matches the structure used internally by the package.
• Added short-term versions of the visitation() function.
• Breaking Added support for setting the initial state in the visitation() function

Performance

• Breaking Removed default naming of cells for samc objects created from rasters. This leads to substantially smaller samc objects, especially as raster inputs become larger.
• Overhauled the samc object creation to be substantially more memory efficient. It is now feasible to create samc objects with 100+ million transient states with 32 GB of RAM. However, this memory efficiency comes with the tradeoff that samc objects can take significantly longer to create (~2x as long based on preliminary testing).
• Added optional support for iterative solvers in metrics (where applicable). This greatly reduces the memory requirements of these metrics, but in general, will take longer to calculate. Initial tests indicate that the visitation() function is feasible for samc objects with 50 Million cells with 32 GB of RAM. Details about changing the solver can be found in the help documentation for the samc-class.
• Added caching behavior for some metrics when using a direct solver. This can reduce run-time by over 95% when rerunning these metrics with the same arguments but different input values. Using different arguments or metrics may require rebuilding the cache, so it is best to keep specific usages of a metric grouped in code.
• In rare cases, cached data is reused for other combinations of arguments and other metrics. This can lead to unexpected situational speedups depending on the order of metrics in the code. Combinations of metrics benefiting from this behavior is not currently documented, but may be in the future.
• Currently, each samc object has it's own cache. Creating multiple samc objects can lead to high memory usage by data caches. The best way to avoid this is to only have one samc object at a time. A future version will implement a single global data cache so that having multiple samc objects will not lead to multiple data caches consuming excess memory.

Website

• Updated the performance vignette to include additional information about the choice of linear solver. Also removed old memory consumption benchmarks due to a flaw in testing where profilers in R do not measure the memory consumption of native code.

Other

• Breaking Combined the three original example data objects into a single list. Updated documentation accordingly.
• Moved the maze example vignette data into a built-in data object.
• Breaking Removed support for TransitionLayer inputs to the samc() function so that gdistance can be removed as a dependency.
• Breaking The sym option for creating the samc object is currently ignored.
• Breaking The map() function was updated so that the output matches the input types used in the samc() function.
• Breaking Rename the tr_args parameter to model to reflect future anticipated support for different types of models. Current usage will not change and assumes a default random-walk model.
• Breaking Renamed the occ parameter in metrics to init (short for "initial state" or "initialize")
• Breaking Added the parameter for setting the initial state in the cond_passage() function to match other metrics, but it is not currently used.
• Bumped various package version requirements.

samc 2.0.1

• Fix debian-clang build error
• Replace built-in progress counter with RcppThread progress counter
• Added/updated maze example vignettes
• Added coin flip vignette

samc 2.0.0

• Removed backward compatibility for deprecated samc() function parameters. This is a breaking change that will make maintaining the package and adding new features a simpler process going forward, and that will hopefully only be a minor inconvenience for users. The warning message on package load introduced in v1.4.0 has been updated to reflect the new changes.
• Updated cond_passage() to return 0 for when i==j in the vectors. This fixes an issue associated with shifted indices in cond_passage(samc, dest). It also technically breaks backward compatibility for when dest equals origin in cond_passage(samc, origin, dest). Previously, cond_passage(samc, origin, dest) would return NA when origin equaled dest, but this decision was arbitrary. The cond_passage() documentation explains why.
• Added a new section for worked examples on the website.
• Added a new example illustrating how to use various aspects of the package with a simple perfect maze and interpret the results. See the Maze Part 1 vignette.
• Added multithreading for the dispersal(samc, origin/occ) function via the RcppThread package. See the Parallel Computing vignette for details.

samc 1.4.1

• Added an input check for multiple absorption that throws a more informative error when a list contains anything other than matrices
• Updated the crs check in samc(). CRS objects have a hidden field that can vary depending on system and software versions, and previous versions of the check would not account for this. This would to lead to false positives where perfectly compatible rasters were reported as incompatible. The corresponding error message was also fixed to report the correct issue; the code was initially copied and modified from another input check, but the error message wasn't updated in the process.
• Added an initial vignette discussing Disconnected Data. The current contents are only slightly modified from an email discussion; they will be rewritten and expanded upon in the future. The Troubleshooting vignette has had an error message and a warning message related to the topic added to it.
• Added a Rcpp related error to the Troubleshooting vignette.
• Bumped version requirements for R to 3.6.0, Rcpp to 1.0.5, RcppEigen to 0.3.3.9.1, and set C++14 as the standard to use in Makevars.
• Enabled the Github discussions page as a replacement for Gitter

samc 1.4.0

• Due to a ballooning parameter count, the samc() function parameters are being adjusted. The new version is samc(data, absorption, fidelity, tr_args). Code using the previous syntax should continue to work (with one rare edge-case as an exception), but backward compatibility will be removed in version 1.5.0, so old code should be updated. See the samc() function documentation and website tutorials for full details and examples. Package startup output has been added to detail the changes as well.
• Updated long/lat handling in samc() to use projection info built into the raster. Deprecated latlon parameter (no longer needed). Added warning for when rasters have non-square cells and are missing projection information.
• The data parameter should be used to pass in the data related to transition probabilities (essentially replaces the resistance and p_mat parameters)
• The tr_fun and directions parameters have been deprecated. This information is now passed as list to the tr_args.
• Deprecated override parameter in the samc() function. See samc-class documentation for details on how to set this.
• Added support for specifying if transition functions are symmetrical or not through the tr_args parameter list.
• Added the ability to directly input a custom TransitionLayer to the samc() function. This allows more flexibility than RasterLayer/matrix maps, but is a little safer than directly inputting a P matrix. See samc() documentation and Overview vignette for full details.
• Added the ability to use the $ operator for accessing and modifying components of samc-class objects. See samc-class documentation for details.
• Updated check() so that multiple rasters can be inputted in the first argument as a RasterStack. This eliminates the need to manually run check() for multiple pairs of rasters.
• Added initial support for caching intermediate results of some calculations. This currently only benefits dispersal(samc, occ), which now caches the diag(F) calculation. This means that while the first run of this method will still be slow, subsequent runs will be substantially faster. With this feature, dispersal(samc, origin) has been enabled and will share the same cached information with dispersal(samc, occ). Future versions will expand the cache options to additional metrics.
• Added support for multiple absorption. The absorption parameter in samc() is treated as the total absorption (consistent with previous behavior). After the creation of the samc-class object, additional absorbing states can be attached to the samc-class object. See the samc-class documentation and the new Multiple Absorption tutorial for more details.
• Added a new absorption() metric. This metric is closely related to the mortality() metric. The absorption() metric can be used to determine the overall probability that a particular absorbing state will be reached (the mortality() metric calculates it for individual transient states rather than overall).
• Fix missing value short-term dispersal
• Overhauled the Overview vignette, including adding more details about the construction of the P matrix.
• Performance vignette update
• Updated documentation for various analytical functions, including more formal/consistent terminology.
• Vector outputs from metric should now all have named cells. These names correspond to the row/column names of the P matrix.

samc 1.3.0

• Fixed an issue with the check() function when data contains NA's.
• Fixed an issue with the raster returned from locate(samc) having 0 for NA cells.
• Improved error checking and messaging for the check() and locate() functions.
• Named rows and columns for the P matrix are now supported. Previously, naming the rows and columns would cause some checks to fail. If names are not manually assigned, the names are simply the row/column numbers converted to character strings.
• Analytical functions updated to support named inputs for the origin and dest location parameters
• When both the origin and dest parameter is used in a function, the inputs can be paired vectors.
• Added the pairwise() utility function
• Created a new Locations tutorial vignette for new location input options.

samc 1.2.1

• Fixed a regression in v1.2.0 where the samc() function would not work correctly unless matrix/raster layers contained at least one NA cell
• Revamped the automated test suite with more test scenarios to better catch issues before release
• Added checks during samc-class creation to prevent potential issues with discontinuous/clumped input data. Currently, this type of data will not work with the cond_passage() function, but will in a future release.
• Reworked some of the vignettes to produce cleaner pages and remove suggested dependencies (e.g. gifski, gganimate, ggplot2) from the package so that users aren't bugged about installing them if they don't need them.

samc 1.2.0

• Added the ability to create samc-class objects from a custom P matrix using p_mat parameter in samc(). See the samc() documentation for details
• Added the cond_passage() function, which calculates conditional mean first passage times
• Added the locate() function, which functions similarly to the cellFromXY() function in the raster package. It's used to get cell numbers from xy coords, but unlike cellFromXY(), it properly accounts for how cells are numbered when the P matrix is constructed.
• Adjusted the absorption inputs to support values of 0 (i.e., no absorption). Currently, at least one cell must have a non-zero value
• Fixed an issue where raster/matrix inputs containing isolated cells (individual cells neighbored by only NA values) would lead to malformed P matrices.

samc 1.1.0

• Added support for the use vectors of time steps in most short-term metrics. It is more computationally efficient and ergonomic to do this rather than calculating short-term metrics for time steps individually. Some key points:
• When vector inputs are used for time steps, the result is contained in a list
• The names of the entries in the list are character versions of the corresponding time step values
• Time step vectors must consist of ordered positive integers with no duplicate values
• Time step vector inputs have not been added for short-term metrics that return dense matrices
• Updated the map() function to support list inputs generated by the short-term metrics. The result is a list of RasterLayers
• Updated the Temporal Analysis and Animations vignettes to incorporate time step vectors
• Created a Gitter community for package support. Gitter badges on the README and home pages can be used to access it.
• Updated the package citation info to refer to Marx et al. (2020, DOI: 10.1111/ecog.04891)

samc 1.0.4

• Add conditional usage of suggested packages in vignettes
• Minor updates for package info

samc 1.0.3

• First CRAN submission

samc 1.0.0

• Complete package rewrite (code dependent on v0.1.0 will not work)
• samc-class for managing SAMC data
• Utility functions for creating samc-class objects, checking inputs, and mapping data
• Heavily optimized analytical functions for all metrics described in Fletcher et al. (2019, DOI: 10.1111/ele.13333)
• Utilizing sparse matrices
• Eigen C++ implementation via Rcpp and RcppEigen
• Updated example data
• Extensive documentation
• Several tutorials

samc 0.1.0

• Created crude functions for calculating metrics in Fletcher et al. (2019, DOI: 10.1111/ele.13333)
• Included example data

andrewmarx/samc documentation built on Nov. 1, 2024, 10:10 p.m.