CVS log for pyrobot/brain/governor.py
	Help

Request diff between arbitrary revisions

Changed internal names of circularlist (next -> nextPos)

Added the GovenorNetwork, and examples

Changes to allow "pattern correct" display.

Working on commenting

Pyro -> Pyrobot renaming

Added method to train from RAVQ buffer without giving a pattern

Changed the behavior of setLearning() in the governor

Added net.shareSomeWeights() in conx

Changed a test call to setLearning(0) to governing = 0

Doug and Lisa worked together.

ravq.py: changed some of the verbosity settings so that you would only
see model vectors at higher levels.

conx.py: added mapInputs and mapTargets

governor.py: Re-did the balanced mask, overrode addLayer,
addContextLayer, and connect.  Re-worked the networkStep method, still
need to write this for the base class.

Governor: removed extra funcs that weren't needed; removed add() for layers, becuase that won't work right in AddThreeLayers (it creates layers that get added to both networks); blob is fine

In GovernorSRN, added override methods for add and connect.  One is
not needed for addContext, since it will call add.

Updated the networkStep method so that it could handle multiple input
and output layers.

Also updated it to check if the trainingNetwork had sharedWeights
turned on, if not, it takes care of it.

Added error checking to make sure that epsilon was not changed before layers are added

Made setEpsilon and setMomentum work with only one parameter (which will be used for both if the second is not given)

updated setMomentum and setEpsilon for all gov networks

changed to allow the "live" network and the "governed" network to have
different epsilons and momentums.

Flush Governor prints too

Made sure CrossValidation does now learn

Set learning appropriately

Fixed a conx bug in setting context activations in SRN.step; updated sequence tests to run with latest conx, and tested GovernorSRN

Added train(cont=1) in conx; add setMaskWeight(1, 2, 1) to governor nets

setLearning overloaded to turn off all bits that might change

Removed callbacks; better decay with own decayHistogram; updated NNPredict brains

Added some pause; got rid of callback in ravq

Made governor report names rather than positions

Changed decay to to remove any m.v. that wasn't used since histogram was reset

conx bug: orderedInputs was not 1 in SRN; added auto mask for gov networks; var watcher -> exp watcher; aibo.py asks for url/ip

updated the mask

Added command-line parameter; turned on governor

Turned momentum off

Made a couple of new flags: useCrossValidationToStop, sweepReportRate; removed sequence2.py

conx: protect divide by zero error; governor made to train data

added a setContext on cross-validation of SRN

Conx: moved Network.step in SRN to method for override; governor now is ready for valid parameters

testing srn gov stuff still in progress

Working on Governed SRN

Primed the RAVQ first, with 25 model vectors, then trained 90% correct

Rearranged a bit; cleaned up output

Brand spanking new! All lists are CircularList; got rid of lots of extra code

Added some test code that can be removed later (self.counts)

ga added pos and popSize; gaca uses those; gov? Activ added try catch to help Tk's smp woes

More comments on weighting.

flipped something by accident in the mask. fixed now.

Governor code now includes example based on real experimental data
pulled from the multiple labelling offline data.

The governor versions from recent experiments. These versions have
been slightly modified to use default values and utilize some
inheritance.

bug fix in winnerCount

Added to governor.

RAVQ governer implementation

Adding AriaService interface for gripper/ptz.

Draft of Governor nnet + som