optimizations.svn-base

Probabilistic graphical models in matlab.

POSSIBLE OPTIMIZATIONS:

* In the Message Passing algorithm, we have to divide by the destination's message. It could be better just not to multiply in the first place. Of course, in the case of the Potential Node, it's not clear how we may do that since it would screw up the product_messages implementation. But for Variable Node, it's easy. Not sure of the speed gains.
-> DONE, for the Variable Node only

* In the Gibbs Sampler, don't recreate the functors each time. Big potential speed gains, I would say.
-> DONE

* Clean up the Message Passing and Gibbs Sampler implementation. It could also save on speed.
-> DONE

* Cache something / reimplement completely the way DiscretePotential works. That is cool because I can keep the same interface and make changes under the hood to the implementation.

* Optimize the sums routines over a DiscretePotential

* Cache a lot of computations in the Tree NP Gibbs, since resampling takes place at every stage. (for example in the backward pass, and maybe in other places as well)

* Write the actual optimizations in the code for the ChainedSingleDiscretePotential, now that the implementation and interface are ready.

POSSIBLE IMPROVEMENTS:

* Change the way Gibbs Sampler works by using a ChainedSingleDiscretePotential! Thus we can benefit from any optimization we make to that class...
-> DONE

* Make many Potential classes derive from an abstract class DiscretePotential

* Clean up particle filters, remove all obsolete functions in Potential and RandomVariable

* Get rid of template arguments in the MessageNode classes