[9a5787]: / NEWS  Maximize  Restore  History

Download this file

520 lines (456 with data), 24.1 kB


Various bug fixes from 0.97


* When a model is dumped to jags.dump, all named nodes in the model are

* Major code reorganization with the functions, distributions and samplers
  now in loadable modules. This has implications for ownership of dynamically
  allocated objects, as objects created in a loadable module must also 
  delete them.  The FuncTab and DistTab classes no longer take ownership
  of functions/distributions they contain.
* The Compiler class uses static tables of distributions and functions (as
  these must be accessible to loadable modules). The Model class uses
  a static vector of sampler factories (for the same reason).
* When reading values from NodeArray or SymTab, a boolean function is
  used to select the required nodes. The functions to read all values are
  redundant - since you can supply a trival function that always returns
  true -  and have been removed.  The Console::dumpState function takes
  three different flags: DUMP_PARAMETERS, DUMP_DATA, and DUMP_ALL. The
  last flag is used after an error.
* The function Node::isStochastic has been renamed Node::isVariable.

* The lex/flex library is no longer linked.  This was unnecessary, and
  caused problems on IRIX building a shared library.


* All aspects of the log normal distribution (density, distribution
  function, quantile function, random samples) were incorrect due to a
  mis-specification of the standard deviation, which was calculated as
  the *square* of 1/tau (where tau is the scale parameter) instead of
  the *square root* .
* The Chi-squared distribution is referred to as "dchisqr" in the BUGS
  language and not "dchisq" as I previously assumed.
* The quantile function for the negative binomial function was incorrect.

* The following distributions have new restrictions on valid parameters
  dbin(p,n)     n must be discrete-valued.

* Removed redundant test for lm in configure.ac, which was causing 
  warnings on IRIX and failures on Solaris.
* Fixed test for sunperf in acx_blas.m4 to use -xlic_lib=sunperf.

* LeftRangeIterator and RightRangeIterator have been replaced
  with a single class (RangeIterator) with member functions
  nextLeft and nextRight. Counter is now a scalar RangeIterator.
* SubSetIndex is now obsolete and has been removed.
* Fixed bugs revealed by the SGI CC compiler. Unfortunately, the MipsPro
  compiler refuses to do dynamic casting to an incomplete type, a 
  technique which was used a lot in the JAGS source. In fact
  this not covered by the C++ standard (although it is supported by the
  GNU compiler) so I have been forced to rewrite some of the Distribution
  and Sampler code.


* Fixed out-of-range error when dumping CODA output from a subset
  of a multivariate node.

* New member functions Range::(left/right)Index which invert
* Samplers are now created in the context of a given graph. So
  many of the functions now take an extra Graph argument.
* Model::initialize is now much more efficient. It calculates
  a sampling graph that excludes all uninformative nodes (i.e. nodes
  with no observed descendants), so the Sampler constructors do not
  have to repeatedly do the same calculation.  (Thanks to Geoff Evans,
  who supplied an example where JAGS was extremely inefficient)
* Node::hasDescendant has been removed as it is no longer used and
  there are no bounds on its complexity.
* MixtureNode has been remodelled so that it no longer requires a
  separate OffsetNode. Hence the OffsetNode and OffsetFactory classes
  have been removed. SubSetIndex has been moved from the graph 
  directory to the compiler directory.
* SymTab now calculates the names of MixtureNodes dynamically instead
  of storing them at compile time. This was very inefficient in models
  with a large number of mixture nodes.
* A new sampler - Censored - samples interval-censored observations.
  It is created by the ConjugateFactory.

* The log likelihood was not correctly calculated for unobserved stochastic 
  nodes with a binomial distribution. This affected the PIGS example.
* The default method for checking the parameter dimensions for real-valued
  distributions (DistReal::checkParameterDim) only checked that the
  _first_ parameter was scalar, e.g. in the expression "x ~ dnorm(mu, tau)",
  tau could have been a vector.

* New distribution "dsum" - the sum of two discrete-valued stochastic nodes.
  This is another observable function, like dinterval.
* New function "rank" for calculating the ranks of the elements of a vector.
  This behaves like the R function rank, rather than the BUGS version,
  returning a vector of ranks.
* A deviance node is now created automatically, unless the user has 
  already defined a variable called "deviance".
* Unbounded discrete-valued distributions (e.g. Poisson) can be sampled
  with the new Discrete Slice Sampler.

* Nodes no longer store their name. Names are generated dynamically when
* Identical Range objects now share their memory. Range is now a wrapper
  class around a new class RangeImp which implements the member functions.
* The NodeArray class is more memory efficient as it no longer uses an
  internal map between Ranges and Nodes.
* Changed names of NodeArray::get and NodeArray::set to "insert" and "find",
  following STL conventions more closely.
* Unused class SubSetNode removed.
* UnaryMinus function renamed to NEG
* The BUGSModel class is now used by the Compiler. It has overloaded methods
  for setting and clearing monitors based on name and range.
* The parameters of a LogicalNode are set in the constructor.
* OffsetNodes are no longer constructed for 1-dimensional mixture models,
  since they are redundant.
* A new Sampler class Slicer implements both the the "doubling" method 
  and "stepping" method for slice sampling. The previous class SliceSampler
  is replaced with RealSliceSampler, which is a subclass of Slicer.  
  DiscreteSliceSampler and DSumSampler are both subclasses.
* Improved performance of BUGSModel::coda by avoiding unneccesary buffer
  flushing from the std::endl manipulater: using '\n' instead.
* Improved performance of ConjugateNormalSampler and ConjugateGammaSampler
  by reducing the number of calls to Sampler::setValue from 3 to 2.


* Constant nodes can no longer be initialized.  
* The command PARAMETERS IN no longer ignores initial values for nodes
  that have been declared but not defined. It will give an error message
* MixtureNodes were sometimes being unnecessarily created when fixed subsets
  were being taken. 

* Variable declaration is now optional. If you omit a variable declaration,
  then JAGS will attempt to determine its dimension from the data table,
  and failing this from the LHS of the relations.
* Data transformations can now be defined inside a separate data block.
  See the manual for details
* In data and parameter files, attributes of R structures are now correctly
  parsed.  You can now dump, for example, matrices with dimnames and use
  them as data or initial values.
* A new function "sort" for sorting the elements of a vector (untested).
* C-style block quotes are now allowed in model files.
* The progress bar printed during updating is longer (40 stars), and a 
  guide bar shows you how long it will be.  The UPDATE command has a BY option
  so that you can alter the refresh rate during updates.

* Compiler changes
  - The compiler classes have been split off into a new directory.
  - All deterministic node subclasses, along with Constant Nodes, have their
    own factory object, which is used to maximize reuse of existing nodes
    when compiling the model. 
  These changes have led to modest speed increases, but there is a dramatic
  improvement for the LEUK example which now runs 2.5 times faster.
* More file input/output moved from the JAGS library into the terminal
* Compiler::fillConstants is now gone.  This role is taken over by
* The above change means that there is no need for a default constructor
  for ConstantNodes, so this is now gone.  All ConstantNodes have their
  value set when they are constructed.
* A new BUGSModel class, which is currently unused.
* The Fortran linking fix for MacOs X has been improved. We no longer
  redifine the official autoconf macros. Thanks to Bill Northcott.

* Modified parser and scanner in the terminal front-end to use C++
  strings instead of C character arrays. This solves a problem on 
  SuSE with variable and file names being mangled.
* Attempting to initialize the model before compiling it no longer leads
  to a crash.
* Parse errors occuring in data and parameter files are now more clearly
* Sampling order has been changed so that a node is always sampled before
  its parents. This appears to help in problems where there is a large
  transient after initialization (birats, kidney).
* Three configuration fixes for Mac Os X: We now have a modified version
  of AC_F77_LIBRARY_LDFLAGS that strips lcrt1.o and lcrt2.o. We link to
  -lcc_dynamic on Mac Os X. We check for BLAS and LAPACK in vecLib
  using a modified version of acx_blas.m4 from Octave. Thanks to Chris
* The class ScalarArgScalarFunc was not necessary and has been removed.
* Samplers may now have a "burnin" mode to adapt their behaviour at the
  beginning of the chain. Burnin mode is turned off when the first monitor
  is set.
* The distribution of a StochasticNode is now set in the constructor.
  The setDistribution member function is removed.
* The function of a LogicalNode is set in the constructor.
  The setFunction member function is removed.
* The compiler has been changed so that expressions "a <- b" are handled
  more efficiently. The identity function is no longer required and has
  been removed. The _fast flag is no longer required in conjugate samplers.
* The algorithm in classifyNode (Sampler.cc) is more efficient as it
  avoids repeat testing of the same node.
* Rmath.h is now systematically included last. This prevents problems with
  symbols being redefined using the C preprocessor.
* R-style comments can now be embedded in data files.

0.25 -> 0.50
* Fixed a bug in DMNorm::randomsample that affected multivariate normal
  variables of length 3 or more, and another bug that gave a length
  mismatch error when forward sampling multivariate normal nodes.
* Fixed a bug in ConjugateMNormal sampler that affected nodes with
  multivariate normal children.
* The SCHOOLS example now runs correctly.
* Created a new class SamplerFactory for creating samplers. The process
  of choosing samplers is now more general and this paves the way for
  samplers that update several nodes at once.
* Version number bumped up to 0.50 to celebrate beta status.

* Replaced the ARM sampler with a slice sampler.
* Fixed a bug in the Conjugate Normal sampler which prevented it from
  recognizing cases where it could sample. This bug had also been copied
  to the multivariate normal sampler.
* The two changes above give some important speed improvements.
* LICENCE CHANGE. The license is now pure GPL, with no exception for linking
  with the arms software by Wally Gilks.
* ConstantNodes are now a distinct class from DeterministicNodes.
* There are now restrictions on Data and Parameter (initial value) files.
  It is illegal to set the value of a logical node, as doing so may leave
  the graph in an inconsistent state (which, incidentally, may put the 
  slice sampler in an infinite loop).  The values of deterministic nodes
  are now calculated internally, but you may set the values of constant
  nodes in the data file, and stochastic nodes in either file.
* Distribution::checkParameterDim now checks the value as well as the
  parameters, like Function::checkParameterDim.
* The abstract class DistScalar has been removed. It was not necessary.
* Fixed a bug in DMulti that only allowed the parameter p to be scalar.
  DInterval should also now work with vector cutpoints.
* Fixed the ConjugateMNormal sampler, which now runs the BIRATS example

* The matrix exponential is now calculated by the standard method of Pade
  approximation with scaling and squaring.
* The ConjugateGamma sampler now works with mixture models. Thanks to
  Bettina Gruen <gruen@ci.tuwien.ac.at> for the patch.
* The "Inits" statement is deprecated, and is replaced by the "parameters in"
  statement, which can be used at any time to set the value of unobserved
  nodes.  New statements "data to" and "parameters to" can be used to
  write the data values, or current parameter values, to file.
* Fixed a bug in Model::checkGraph which prevented forward sampling of
  non-informative Poisson. Thanks to Greg Ridgeway <gregr@rand.org>.
* The ParseTree class has been rewritten, with consequent changes to
  the parser and compiler.
* Some minor optimizations, e.g. SArray::value is inlined.
* Functions with a variable number of arguments are now possible.
  The functions min and max, and the inline functions '*' and '+' 
  now have variable arguments. This leads to a reduction in the number
  of nodes in the graph and a consequent increase in speed (along with
  the optimizations).  The EPIL example now runs about 25% faster.
* Min and Max now act like the equivalent functions in R: they
  take a variable number of array arguments, and will return the
  maximum of all the values present in those arguments. This means
  that the incompatibility with BUGS introduced in version 0.21 is
* Fixed a bug in forward sampling of truncated discrete distributions
  and truncated binomial distributions, in which the lower limit
  was never sampled. Thanks to Greg Ridgeway <gregr@rand.org>

* Fixed the ConjugateMNormal sampler so that it now handles more cases.

* max() and min() functions now take a single vector argument, like mean()
* Seed setting fixed in the JAGS terminal. Previously the seed command
  did nothing.  Thanks to Bettina Gruen <gruen@ci.tuwien.ac.at>
* JAGS can now be built in a separate directory from the source directory.
  In particular, "make distcheck" now works.
* Fixed error introduced in 0.20 when calculating upper bounds.
* Responsibility for updating noninformative nodes has
  shifted from the Sampler to the Model. This means that these nodes are
  updated only once per iteration, with consequent gains in efficiency.
* The ForwardSampler class is no longer necessary.
* Graph::getSortedNodes has more built-in checks, and will throw an
  exception if they fail.
* There was a serious bug in DistDiscrete::randomSample leading to 
  incorrect samples being drawn. This has been fixed.
* Conjugate samplers now work correctly with truncated variables.
* The dump format has changed. It now uses the same R dump format as
  initial value and data files.
* MATHLIB_STANDALONE is now defined in the header file config.h
* ARMSampler now sticks at the current value if the loglikelihood is 
  infinite. This occurs in the LITTERS example.
* Model::initialize no longer enters an infinite loop if an uninitialized
  constant node is used in the model.
* The inverse link functions ILogit and Phi could return a value
  of exactly 1 due to rounding error, which would lead to some nodes being
  apparently inconsistent with the data.  These functions now return 1 -
  DBL_EPSILON instead. Similarly they return DBL_EPSILON instead of 0.

* Changes required to build JAGS with gcc 2.95-4 and 3.0.4
  Thanks to Chris Jackson <chris.jackson@imperial.ac.uk>.
* Installation of the shared library has been disabled by
  default. You can re-enable it with ./configure --enable-shared.
* Rearranged headers to minimize dependencies in the JAGS library.
* DMulti is no longer (and should never have been) a subclass of DistScalar.
* Try to speed up calculation of binomial likelihood when size parameter
  is constant.
* Corrected random samplers for Dirichlet and Discrete distributions.
* ForwardSampler now updates all deterministic children of sampled node.
* Fixed constructor for Model (_cansample is now initialized to false).

* Bugfix release.  Scalar nodes were causing a compiler error.

* Separate installation of terminal and library.

* Corrected CODA output with thinning interval != 1
* Added Bernoulli distribution to DistTab (It had been forgotten)
* Allow negative numbers as parameters to distributions and the
  truncation construct T(,).
* Added separate initialize command to scripting language (this is
  required to handle multiple chains, which we can't do yet, but
  it is better to put this in place now)
* Fixed bug in CODA output which claimed that whole nodes were
  not being monitored

* Allow C-style block comments in script files
* Use config.h
* Use T(,) syntax for truncation to avoid confusion with I(,)
  in WinBUGS.

* Changes to the console, now renamed the terminal.  Now uses a Stata
  style scripting language.
* Distribution class names and file names have been changed to 
  correspond with the names in the BUGS language
* DistTab is a proper class. You can now add new distributions to
  the DistTab.
* FuncTab is a proper class, like DistTab
* More extensive use of Standard Template Library
* Lots of documentation
* Lots of cruft removal

* SubSetNodes with variable indices have been removed completely
  (were deprecated in release 0.12)
* NodeArray has been taken out of the graph directory and put in
  the compiler directory

No major changes in functionality. Some architecture changes instead
* GraphNode is renamed to Node
* Parameter is renamed to SArray (a more sensible name, given that it
  is based on arrays in S)
* Node is no longer a subclass of SArray. Instead, Node contains an
  SArray as a public data member.
* Major changes in the directory structure.  Source for the console is
  now in a separate directory from the JAGS library, which is in turn
  divided into several subdirectories. Currently, the JAGS library
  is a "convenience library" created with libtool, so is not installed.
  This will change when the class interfaces are nailed down.


* ConjugateGamma sampler now works with Weibull children
  and Double exponential children.
* GraphNode::hasDescendant fixed (would incorrectly return false
  under certain circumstances).
* FiniteSampler fixed (would always sample from the prior).
* Conjugate Normal sampler now works correctly with bounded nodes.
* New distribution "dinterval" (class DistInterval) for handling
  interval censored observations.
* Conjugate Normal sampler now works correctly with mixture models.
* Added matrix functions inverse and logdet.
* Added Wishart and Multivariate Normal distributions.
* Added Conjugate samplers for Wishart and Multivariate Normal
* SubSetNodes with variable indices are now deprecated: Mixture
  models are now represented with a new class "MixtureNode",
  and related class "OffsetNode".


* Added conjugate samplers for normal, gamma (including exponential and
  chi square) and beta distributions.
* Fixed important bug that led to likelihood being evaluated twice (!)
* ATTENTION: The dump format has changed for Trace Monitors.


* New experimental distribution dmstate for transitions in multi-state
  Markov models.
* BLAS and LAPACK libraries now required (for the above, but they
  will be needed for multivariate normal and Wishart distributions later)
* Parameter class is no longer a virtual class. Takes some functionality
  from GraphNode and avoids a lot of code repetition. However, there
  may be a performance hit.
* Parameter checking added for functions and distributions.
* addParent and removeParent are no longer virtual functions (a preliminary
  for removing graphical computations to a separate library)


* Further speed improvements. The parser tries to interpret index
  expressions on the right hand side of a relation in order to avoid
  creating unnecessary nodes and - in particular - SubSet nodes, which
  are computationally very expensive.  There is a corresponding increase
  in speed for the BONES example.


* The ArrayNode class is abolished and is replaced with two different
  NodeArray: A container for nodes. Also a factory object that will
             create aggregate nodes or subset nodes to represent
             fixed/stochastic subsets respectively.
  AggNode:   Aggregate nodes. Simply copies its values from other
             nodes. Several different nodes can be aggregated together
             to create an AggNode.

  This leads to an enormous improvement in speed for the LSAT example.
* Runs the KIDNEY example


* Updating of deterministic nodes no longer takes place automatically,
  but is the responsibility of the sampler. This was done in response
  to profiling of the LSAT example.


* Nested indexing now works


* Fixed segfault when quitting before defining a model.
* User may now choose the thinning interval for monitors
* Runs the LITTERS example
* funclib remodelled to use the new Parameter base class.
  FuncTrees are now obsolete. The compiler creates additional
  logical nodes for complex expressions.
  NOTE: This change creates an enormous performance hit in the
  EPIL example, which has a rather complex linear expression. This
  suggests that a special function for bi-linear expressions 
  may be useful in a future version.
* Added regression test suite based on the classic BUGS examples, vol1.
  Since this is based on copyright material, it is not included in
  the distribution.


* Major rewrite of libdist with corresponding changes to Stochastic nodes
* Introduced a new base class Parameter
* Some examples broken:
  - epil2.bug crashes with a hard-to-trace bug that is overwriting memory
  - blockert.bug crashes with error from arms sampler
  Remember this is still alpha software


* License conditions on ARMS software clarified in file AUTHORS
* Runs the PUMP example
* Bug in logDensity for the Gamma distribution fixed
* Monitoring of whole multivariate nodes now works correctly
* Names of nodes in dumped output from TraceMonitor are now correct
  (when  monitoring the whole node)
* Runs the SEEDS example (seeds.bug, seedssig.bug seedsuni.bug, seedspar.bug)
* Fixed updating of array nodes by adding the contents as parents
* Runs the SURGICAL example (surgical.bug, surgfix.bug, surgrnk.bug)
* Changed ARMSsampler so that it starts with extremely conservative
  bounds around the current value.
* Runs the SALM example (with gamma and pareto prior)
* Runs the DYES example
* Runs the STACKS example (with normal, logistic, double exponential, and
  t_4 errors)
* Runs the EPIL example (epil2.bug, epil3.bug)
* Runs the BLOCKER example (blocker.bug, blockert.bug, but NOT blockht.bug)
* Runs the OXFORD example (but with poor agreement for sigma due to
  poor mixing of this parameter)
* Runs the LITTERS example


* Fixed bugs in TraceMonitor
  - Last monitored value correctly recorded
  - mcpar attributes now correct
* Added initialize method for Model
* Nodes now sampled in forward-sampling order (but is this efficient?)
* Initialize includes check that graph is complete
* Can now monitor (some) multivariate nodes
* Dump format changed. It is no longer compatible with the mcmc
  class used in the coda library, so the class attribute is now
  "mcarray".  This is a provisional name
* Runs the first real example: RATS, including the version
  with missing values (data in "ratsmiss.dat"), but
  - very slow because there is no conjugate sampler, and
  - higher autocorrelation than BUGS. This may be due to the 
    choice of sampling order.


First public release (18/12/2002). Runs the "line" example, but does
little else.