Buchtitel: Proceedings of Intelligent Adaptive Systems Workshop IAS-95, Melbourne-beach, Florida
When applying Inductive Logic Programming techniques in real-world settings, many problems will come up. Since not many real-world applications are reported using ILP-techniques this report describes some problems that may appear when starting such a project. The project was done in co-operation with Högdalenverket, a heat and power plant burning household refuse in the Stockholm area, Sweden. The application problems with collecting data and the application of ILP-techniques are discussed. Results of tests performed while using SPECTRE, an ILP-algorithm developed at Stockholm University, are reported. These results show that the addition of background knowledge and addition/retraction of parameters has positive effect on the performance of the ILP-techniques. After initial tests a knowledge acquisition stage was started. This resulted in knowledge about the domain that was used for evaluating the results of learning using SPECTRE. This prevented several mistakes by interpreting data and evaluating performance. This knowledge was partly used as background knowledge for the learning algorithm. Noise handling was applied and increased the efficiency of the SPECTRE-algorithm. The paper lists the results of tests performed on data from the refuse-burning plant. It also compares the performance of the SPECTRE-algorithm on theoretical databases with the performance on this practical domain. It also gives some insight in the change of performance that can be expected when transferring from the laboratory-domains to domains gathered in the real-world.