With the move to a new web server, we had the opportunity to give the website a bit of a face lift. In this first phase, the only noticeable (hopefully) difference will be in some minor css cosmetics. In the background, however, the entire system has moved to a dynamic PHP / MySQL driven platform. It is nice to have entered the 21st century.

New features are on there way down the pipeline, but until then the most noticeable difference should be much more frequent updates, thanks to the new web admin interface. So stay tuned...

Results showing the effects of using a threshold classification rule as opposed to a majority wins rule are now available. Thresholding appears to be effective in shifting errors towards those that exclude biological sweeps as opposed to including non-biological sweeps. Timing Benchmarks are also available. These benchmarks indicate that an overwhelming majority of experiment time is spent training the classifier.

Results showing the effects of varying the momentum between 0.1 and 0.9 are now available. Initial results indicate that significant time savings may not be achievable by using a different momentum term.

A number of optimization experiments have been performed using a multilayer perceptron (a form of neural network). The experiments include optimization various parameters including structure, learning rate, and decay. An attribute subset experiment was also performed with the goal of finding a subset of attributes that reduce the training time of the classifier without substantially reducing classification accuracy.

A comprehensive subset experiment using the Naive Bayes classifier has been performed and the results are available in the experiments section.

The attribute optimization tests have been completed. Initial tests show that an attribute combination of Reflectivity, Reflectivity Skewness, Spectrum Width, Radial Velocity Skewness, Range, Radial Velocity, and Reflectivity Kurtosis, produce the best results. Hold one out validation with 16 training files yielded an accuracy of 93.75%. The experiments used a 25x25 second order moment window. See the results here.