Engineering topics
Publications and source records attributed to Kanefsky, Bob.
Report describes continuing development of software for constraint-based scheduling system implemented eventually on massively parallel computer. Based on machine learning as means of improving scheduling. Designed to learn when to change search strategy by analyzing search progress and learning general conditions under which resource bottleneck occurs.
The explosive growth of the World-Wide Web (WWW) in the past year has made it feasible to provide interactive graphical tools to assist scientists in locating planetary images. The highest available resolution images of any site of interest can be quickly found on a map or plot, and, if online, displayed immediately on nearly any computer equipped with a color screen, an Internet connection, and any of the free WWW browsers. The same tools may also be of interest to educators, students, and the general public. Image finding tools have been implemented covering most of the solar system: Earth, Mars, and the moons and planets imaged by Voyager. The Mars image-finder, which plots the footprints of all the high-resolution Viking Orbiter images and can be used to display any that are available online, also contains a complete scrollable atlas and hypertext gazetteer to help locating areas. The Earth image-finder is linked to thousands of Shuttle images stored at NASA/JSC, and displays them as red dots on a globe. The Voyager image-finder plots images as dots, by longitude and apparent target size, linked to online images. The locator (URL) for the top-level page is http: //ic-www.arc.nasa.gov/ic/projects/bayes-group/Atlas/. Through the efforts of the Planetary Data System and other organizations, hundreds of thousands of planetary images are now available on CD-ROM, and many of these have been made available on the WWW. However, locating images of a desired site is still problematic, in practice. For example, many scientists studying Mars use digital image maps, which are one third the resolution of Viking Orbiter survey images. When they douse Viking Orbiter images, they often work with photographically printed hardcopies, which lack the flexibility of digital images: magnification, contrast stretching, and other basic image-processing techniques offered by off-the-shelf software. From the perspective of someone working on an experimental image processing technique for super-resolution, the discovery that potential users are often not using the highest resolution already available, nor using conventional image processing techniques, was surprising. This motivated the present work.
This research note shows the results of applying a new massively parallel version of the automatic classification program (AutoClass IV) to a particular Landsat/TM image. The previous results for this image were produced using a "subsampling" technique because of the image size. The new massively parallel version of AutoClass allows the complete image to be classified without "subsampling", thus yielding improved results. The area in question is the FIFE study area in Kansas, and the classes AutoClass found show many interesting subtle variations in types of ground cover. Displays of the spatial distributions of these classes make up the bulk of this report. While the spatial distribution of some of these classes make their interpretation easy, most of the classes require detailed knowledge of the area for their full interpretation. We hope that some who receive this document can help us in understanding these classes. One of the motivations of this exercise was to test the new version of AutoClass (IV) that allows for correlation among the variables within a class. The scatter plots associated with the classes show that this correlation information is important in separating the classes. The fact that the spatial distribution of each of these classes is far from uniform, even though AutoClass was not given information about positions of pixels, shows that the classes are due to real differences in the image.
It is described how Minimum Description Length (MDL) can be applied to the problem of DNA and protein evolutionary tree reconstruction. If there is a set of mutations that transform a common ancestor into a set of the known sequences, and this description is shorter than the information to encode the known sequences directly, then strong evidence for an evolutionary relationship has been found. A heuristic algorithm is described that searches for the simplest tree (smallest MDL) that finds close to optimal trees on the test data. Various ways of extending the MDL theory to more complex evolutionary relationships are discussed.