Trajan

survey performed by Juha Ikonen, May 21st 1999

Trajan 4.0 is a fully-featured Neural Network simulation package. It includes support for a wide range of Neural Network types, training algorithms, and graphical and statistical feedback on Neural Network performance.

The SOM (or Kohonen Network) is only one of the supported network types and the approach is somewhat different from other applications implementing the SOM. The Kohonen Network is treated as one special case of neural networks in general and some of the special features of SOM have not been implemented. Major defincies include

map topology and shape are always square
some of the useful visualization tools are not included such as U-matrix and trajectory of BMUs

On the other hand many features have been implemented very well, these include preprocessing tools and labelling. Also the speed of training is worth mentioning. In general Trajan is a good package for various purposes and if one does not need advanced visualization tools, Trajan is well suitable also for SOM applications.

Disclaimer: If any information on this page simply is not true, please tell us about it and we'll correct it ASAP.

Disclaimer: The opinions and observation herein should be considered personal of the person having performed by the survey, at the time of the survey. They do not reflect any official standing of his employer, of the Laboratory of Computer and Information Science or the Neural Networks Research Center.

General

Program name	Trajan Neural Networks Simulator, Release 4.0 A
Availability	Commercial software, demonstration version is available at http://www.trajan-software.demon.co.uk. Pricing: `Standard Academic USA, Canada and Mexico: 795 595 (US dollars) Other countries: 595 450 (UK pounds sterling)` Company information: Trajan Software Ltd. Trajan House, 68 Lesbury Close, Chester-le-Street, Co. Durham, DH2 3SR United Kingdom tel/fax: +44 (191) 388 5737 email: andrew@trajan-software.demon.co.uk.
Purpose	Fully-featured Neural Network simulation package for variety of purposes.
Operating system	32-bit Windows (Windows 95/98, Windows NT)
User interface	Easy to use graphical user interface.
Documentation	Extensive online help and a 350-page user manual.
Other	Trajan's features can be used also programmatically via an Application Programming Interface (API), application samples are included.

SOM features

map parameters
Teaching algorithm	Standard SOM algorithm, implementation seems ok.
Map size	2-dimensional map lattice. Smallest 1x1, maximum size is limited by the amount of available memory.
Map lattice and shape	Both rectangular
Neighborhood function	Function type: square
	Neighborhood size (h): decreases linearly from the start to end values during the training.
	Learning rate (alpha): is altered linearly from the first to last epochs.
Initialization	By default the map weights are treated as vectors and set to unit length. Other available methods are - uniform within a range of minimum and maximum values, - Gaussian with mean and standard deviation and - zero, the weights are all set to zero.
Distance function	Probably euclidian, not verified.
Unknown components	Allowed
Teaching length	In epochs, stopping conditions may also be specified. Additionally, if over-learning occurs, the best network discovered during the training can be retrieved.
efficiency
Speed [Windows NT 4.0, 333 MHz Pentium II, 128 MB RAM]	for ~3000 13-dim data samples, 10 epochs training time: 7 seconds
Results	seemed ok quantization error: 0.3358

[Comments on SOM implementation]

Usability

preprocessing
Input formats	Variety of standard ascii formats (tab/comma/space separated), STATISTICA and Trajan's own Fast binary data file format.
Data handling and selection	Very good tools for data preprocessing: - a special data set editor, where user can edit data and select which components and data vectors to use, - data can be scaled using several methods, - missing values can be substituted by mean, median, minimum, maximum or zero value, - nominal (non numeric) variables can be substituted by numerical values, - useless features (vector components) can be ignored from input data set automatically.
postprocessing
Output formats	Variety of standard ascii formats (tab/comma/space separated), STATISTICA and Trajan's own Fast binary data file format.
Map measures	Quantization error (RMS)
Labeling	Advanced: labels can be set by user or automatically
Clustering	By visualization
visualization
Inspection of neurons	Simple: weight vectors can be inspected, no graphical representation available.
Clusters/map shape	Topological map
Correlations	By inspecting weight vectors
Data projections	BMU's can be found for a single vector or for entire test set.
Markers	Labels

http://www.cis.hut.fi/projects/somtoolbox/links/trajan.shtml
somtlbx@mail.cis.hut.fi
Monday, 09-Oct-2000 12:53:09 EEST