- Ease of use - Control - Management - Sophistication - Embedding - Back to Products -
One size suits none
Neural network design may be approached at a number of levels, from the "hands-off" user wanting an automated procedure with good results and minimal demands on expertize and valuable time, through to the expert wanting to determine precise details of the best architecture and learning algorithm for the task at hand. Trajan is carefully designed to support all levels of user. New users can achieve results quickly and easily with Trajan's highly automated Intelligent Problem Solver, while even the most experienced neural network designer will find ample support in the deeper capabilities of the Custom Network Designer - and still find the Intelligent Problem Solver an exceptionally useful tool for exploratory analysis.
One step training
Using the Custom Network Designer, you can combine the most detailed level of control over network design and training, with a "one step" design that covers the entire training process. Training a neural network may include the following steps: data pre-processing; initialization of the network; selection of data subsets for training, early stopping and performance evaluation; selection of regularization coeffcients; execution of the actual training algorithm, and designation of stopping conditions to terminate training; setting of confidence thresholds to optimize performance in classification; pruning of input variables and/or hidden units. In Trajan, all the options relevant to a network type are organized on pages on the training dialog, making detailed customization simple. Default options are designed in a "whole process" mode - for example, SOFM and MLP training dialogs by default initiate an effective two-stage training process, and the RBF dialog a three-stage process. At the same time, via Trajan's Custom training dialogs, you can train hybrid networks using any feasible mixture of algorithms - MLPs with Principal Components optimization of the first layer; RBFs with SOFM or LVQ training of the hidden layer, etc.
Trajan's Network Editor gives you the finest possible level of control. You may individually select a range of pre-processing and post-processing options (including conversion functions for both numeric and nominal variables), choose activation functions and error functions, view weights histograms, and even edit individual weights. You can also route the Custom Network Designer through the Network Editor to perform detailed customization prior to network training.