- Ease of use - Control - Management - Sophistication - Embedding - Back to Products -

Integrated handling of data, models and results
Neural network design requires the generation and comparison of multiple networks. Trajan has unrivalled facilities to support the experimental process. Trajan automatically associates a network file and results file with each data file - open a data file, and Trajan opens the corresponding network and results files too. Network files contain multiple models (i.e. neural networks and network ensembles). You can easily view network summary details, and select a model or models for testing. Results files are hierarchically organized, with the capacity to hold an unlimited number of spreadsheets and graphs. As you perform analyses in Trajan, results spreadsheets and graphs are directed into the results viewer, where you can quickly page through them, then group, arrange and retitle them.

Comparative results generation
Trajan's results generating facilities include performance summary statistics (correlations, confusion matrices, ROC curves), model predictions, response graphs and surfaces, user-defined (one-off) case execution, scatter diagrams, histograms, sensitivity analyses and weight summaries. Wherever appropriate, you can select multiple models and generate comparative results - for example, to compare the predictions of several models side by side in a single spreadsheet, or to plot the response curves of several networks against the same response variable.

Multiple model generation
Trajan's major tools for neural network creation - the Intelligent Problem Solver and the Custom Network Designer - can both generate multiple neural networks; the former, by searching across possible configurations; the latter, by resampling the data set. In either case, the multiple models generated are automatically selected for results generation, allowing immediate comparative analysis.

Network ensembles
Many neural network theorists recommend the use of neural network ensembles, that generate predictions by averaging or voting among the outputs of multiple networks. Trajan has facilities to generate and execute ensembles. Besides the capabilities they have in their own right, Trajan's ensembles make it easy to generate statistics on average performance of networks in an experimental procedure.