New Features of SNNSv4.1

• New batch execution language batchman
• Integration of the genetic algorithm tool Enzo. This tool allows the evolutionary optimization of neural networks. Enzo is available as seperate tar file on our ftp-server at the same location as SNNS. For all enquiries about Enzo please contact enzo-request@ira.uka.de.
• Improved window handling. On re-opening of a SNNS window it gets put on top, instead of issuing an error message.
• New learning algorithm 'scaled conjugated gradient' (SCG)
• The parameter lists in the control panel adapt to the number of paramters required, i.e. all widgets present need to be filled out.
• WinSNNS, a MS-Windows front-end to SNNS batch execution on unix workstations. WinSNNS is NOT part of our official SNNS distribution! It can be found at http://www.lans.ece.utexas.edu/winsnns.html. Please contact ghosh@pine.ece.utexas.edu for more information.
• Optional parameter `Teacher-Forcing' for Jordan networks introduced.
• Validation now possible for BPTT networks
• Extensive debugging:
  • Seg-fault when trying to train a network without output units
  • Seg-fault when pressing the info button without loaded patterns
  • Wrong feedback loops in associative memory nets deleted
  • Wrong activation functions in associative memory when constructed with bignet
  • Wrong initialization function in kohonen networks when constructed with bignet
  • Wrong error computation in Hebbian learning
  • Wrong error computation of analyze function band
  • Wrong MSE computation for validation
  • Better font size adaption of label widgets
  • Pruning now allows validation with 0 cycles
  • GRAPH panel could sometimes not be closed
  • Wrong scanning of update parameters in snnsbat
  • Error in the manual: Missing parameter of BackpropWeightDecay

Last modified: mamier@vasarely Fri Dec 1 00:00:00 MET 1995