Abstract

Prediction of various values is a hot item for many companies. If values such as the volume of energy consumption, or the number of sales of a product can be predicted with good accuracy, then the efficiency of the company can be increased. Especially in this period of recession, efficiency can improve the chance of survival. There is a huge variety of techniques to produce predictions. The optimal technique is ultimately determined by the problem at hand. To fine-tune the solution, a lot of smaller choices have to be made to find the best parameters.
The goal of this project is to offer researchers a computer program that can be used to optimize the techniques that create the predictions. This program should allow them to make changes in the parameters fast, create big tests, and visualise the results of these tests.
The program is implemented in Java as a web service. This way, multiple users can work with the system at the same time, using a web browser. Tests are distributed over multiple computers (by using JINI/JavaSpaces), to reduce the time one has to wait for the results.
The result is a program that can be used much like a web site. The user can insert the problem as a project and then create an outline for the tests. This outline is created by selecting building blocks that modify the data. These blocks can be stacked to create different results.
The results of different tests are gathered automatically, and the user can look at this data from different angles. Using these views of the results, the user can continue the search for a better prediction.
The use of this program offers multiple advantages. The first is that the creation of tests is easier. Instead of manually making changes or writing tools to do so, the user adds a module to the structure. This also maintains the complete picture of the test. The second advantage is that the system uses multiple computers to perform the time-consuming steps. This increases the number of tests that can be done in the same amount of time. The third advantage is that the program is extendible. New modifications or prediction techniques can be added fairly easily, and these new blocks can cooperate with all existing parts to create new predictions.