developmentNotes.md
In blockDemac: parallel DEMC with several parameter blocks

Development Notes for package blockDemac

The bottom layer supports a Differentail Evolution Marcov Chain (DEMC) with prescribed temperature for all density components.

For documentation of usage start with vignettes (SimpleSingleModel.html, ...)

Code is organized by S4 classes. See ClassDiagram.graphml

Entry object is a PopulationSampler. It contains several other classes for specific tasks. JumpProposer: proposed new Jumps based on the current samples AcceptanceTracker: records acceptance rates over some window and calculates acceptance rates for populations SampleLogs: Records sample results and provides access to it (see below) ChainSampler: performs the sampling

The PopulationSampler takes care of distributing the updates of chains.

The ChainSampler does the sampling for one chain. It must be provided with information on each step (setRangeSpecs,StepInfo). It controls a ChainState with all information on current state of the chain. And it lets BlockUpdaters update this ChainState.

The BlockUpdaters object manages all the dependencies among the blocks and intermediate results. It provides access to the BlockUpdaters. It also implements the loop of a single generation, where all BlockUpdaters are invoked to update their corresponding part in the ChainState. Meta-Information on the different blocks are provided by the contained BlockDimensions object.

The update of parameters is performed by a BlockUpdater object. The specific mechanism (updateBlockInChainState) how this is achieved varies by the specific subClass.

The MetropolisBlockUpdater is central to this framework. It calculates an unscaled posterior probability of the current parameter vector and an alternative location after a jump in parameter space. Based on a Metroplis-Decision, the new parameter vector is accepted or rejected.

The FunctionBasedBlockUpdater provides a easy way to implement new Updaters. See File invChiSquareBlockUpdater.R for an example of implementing a Gibbs sampler based on specific statistics and densities.

The update of intermediate results is performed by an IntermediateUpdater.

The setup of both kinds of updaters is helped by BockSpecification and IntermediateSpecification class respectively. These Specifications are usually created using the functions bockSpec and intermediateSpec.

The parallelisation is not done by chains instead of populations. In this way, nChainPop times nPop cores can be used. For an overview see graphics ClusterControlFlow.