- Ease of use - Control - Management - Sophistication - Embedding - Back to Products -
|Ease of use|
Years of experience and user feedback have gone into the development of the most efficient neural development tool available. Trajan is streamlined to support fast exploratory analysis, automated design, and simple generation, organization and display of results. Procedures are organized into "wizard-like" analyses, leading you step by step through the major analysis stages. Extensive use of sophisticated multiple page dialogs gives you quick access to the most commonly required features, while keeping even the most sophisticated options only a few clicks away.
Intelligent Problem Solver
One of Trajan's major features, the Intelligent Problem Solver streamlines the entire neural design process. To run a basic analysis you need do no more than select the variables of interest, and the amount of time to be spent in designing a solution. The Intelligent Problem Solvers does the rest! Exploratory selection of model type and complexity, selection of variables, and training procedures - are all handled by the Intelligent Problem Solver. Multiple models are generated, and the results presented in comparative form - no need for you to decide whether MLPs or RBFs, Back Propagation or Conjugate Gradient Descent are the best approach to training.
As your expertize grows, and you wish for finer control over the design process, it is optionally available - constrain the model type, complexity (hidden units), training time, and data sampling strategy in detail. Even for advanced users wanting to design the fine details of neural architecture, the Intelligent Problem Solver is an invaluable tool for automating exploratory analysis.
Custom Network Designer
The Custom Network Designer streamlines the design and testing of neural networks using specific architectures and algorithms. Choose the network type, number of layers, units, training algorithm and control parameters. For iterative algorithms, you may optionally display the progress of the error function in real time on a line graph. The Custom Network Designer integrates the training process for each network type on a single dialog, including network initialization and "tidyup" processes such as automatic optimization of decision thresholds for classification networks, pruning of unneeded units in conjunction with regularization procedures and generation of performance summary statistics. The Custom Network Designer also supports the conduct of multiple experiments and data resampling techniques such as bootstrapping and cross validation. For really fine-grained control, you can select individual variable conversion functions, activation functions, and error functions. As with the Intelligent Problem Solver, common options are immediately available on Quick tab pages, while detailed customization is only a few clicks away through advanced tab pages.
Neural network analysis inevitably requires the generation of multiple networks, perhaps of different types, architectures, and trained with different samples of the data set. Trajan stores multiple networks - and network ensembles - in a single network file. Networks are listed together with summary statistics, and multiple networks can easily be selected for comparative analysis. Since its possible to generate very large numbers of networks during exploratory analysis, Trajan includes sophisticated support tools to limit the number of networks in a file, and to select networks for deletion according to criteria such as network performance, and the maintenance of diversity.
Graphs and Spreadsheets
Trajan generates numerous graphs and spreadsheets to summarize or probe neural network performance. Most noteworthy is Trajan's unique capability to generate comparative results - multiple networks and network ensembles may be simultaneously selected, and integrated results generated at a click of a button. For example, you can generate a spreadsheet listing the predictions of a number of models on a data set, with a column for each network; or, a response graph showing the predictions of multiple networks as a single input variable is varied, with the prediction of each model represented by a separate line.
Trajan's spreadsheets and graphs have the full range of facilities you would expect: on spreadsheets, customizable precision for numeric variables, drop-down list selection for nominal variables, print facilities, cut and paste support, and export to text files; on graphs, extensive property editing and customization, including line styles, viewing angles and positions, three-dimensional effects, label fonts, colors, axes and legend configuration, and print facilities.
Results are also elegantly organized in Trajan's hierarchical results files, allowing related graphs and spreadsheets to be grouped together. The viewing area can be divided into panes, with a single results spreadsheet or graph in each pane, making comparison and cross-referencing doubly-easy.