A classic problem from chemistry is used to test a conjecture that in domains for which data are most naturally represented by graphs, theories constructed with Inductive Logic Programming (ILP) will significantly outperform those using simpler feature-based methods. One area that has long been associated with graph-based or structural representation and reasoning is organic chemistry. In this field, we consider the problem of predicting the mutagenic activity of small molecules: a property that is related to carcinogenicity, and an important consideration in developing less hazardous drugs. By providing an ILP system with progressively more structural information concerning the molecules, we compare the predictive power of the logical theories constructed against benchmarks set by regression, neural, and tree-based methods. 1 Introduction Constructing theories to explain observations occupies much of the creative hours of scientists and engineers. Programs from the field of Inductiv...

This thesis is about inductive learning, or learning from examples. The goal has been to investigate ways of improving learning algorithms. The chess end-game "King and Rook against King" (KRK) was chosen, and a number of benchmark learning tasks were defined within this domain, sufficient to over-challenge stateof -the-art learning algorithms. The tasks comprised learning rules to distinguish (1) illegal positions and (2) legal positions won optimally in a fixed number of moves. From our experimental results with task (1) the best-performing algorithm was selected and a number of improvements were made. The principal extension to this generalisation method was to alter its representation from classical logic to a non-monotonic formalism. A novel algorithm was developed in this framework to implement rule specialisation, relying on the invention of new predicates. When experimentally tested this combined approach did not at first deliver the expected performance gains due to restrictio...

Relational learning algorithms are of special interest to members of the machine learning community; they offer practical methods for extending the representations used in algorithms that solve supervised learning tasks. Five approaches are currently being explored to address issues involved with using relational representations. This paper surveys algorithms embodying these approaches, summarizes their empirical evaluations, highlights their commonalities, and suggests potential directions for future research. Keywords: supervised learning, representation, relational learning 1 Introduction Relational learning algorithms extend the capabilities of propositional or monadic supervised learning algorithms. Supervised learning algorithms input a set of instances, which are described by a set of predictor descriptors and a target descriptor. These algorithms construct a function (i.e., a concept description) that can predict an instance's target descriptor value given its predictor desc...

First generation expert systems rely on the use of surface knowledge, such as associational or heuristic. This knowledge is typically acquired from domain experts through exhaustive knowledge engineering sessions. On the other hand, second generation knowledge acquisition technology is characterized by two main features: the use of deep knowledge and machine learning. In the paper we review three second generation methods that partially automate the knowledge acquisition process: inductive learning of rules from examples, model-based rule learning, and qualitative model acquisition. Results of their application to some medical domains are presented. Finally, we outline different stages of expert system development. An extended expert system shell schema is presented which includes a knowledge acquisition and a knowledge explanation module. 1 Introduction Knowledge acquisition is a field of artificial intelligence concerned with the development of methods, techniques and tools for buil...

This report describes experiments in learning models for basic flight manoeuvres from behavioural traces of a human pilot when using a flight simulator. A first set of experiments using decision trees is presented. The auto-pilot built with the generated decision trees flies more smoothly than the human pilot. However the results show also that propositional logic-level representations, like decision trees, are inadequate to fully solve the problem. A learning system using a first-order representation is required. However, current Inductive Logic Programming systems have severe limitations when dealing with such complex domains due to inefficiencies of searching large hypothesis spaces. An important issue to make the hypothesis space search tractable and efficient is the use of background knowledge. Some first results are reported based on a system under development that already shows some uses of background knowledge at a "local" level of learning a single predicate. Identification of...

Most of KDD applications consider databases as static objects, and however many databases are inherently temporal, i.e., they store the evolution of each object with the passage of time. Thus, regularities about the dynamics of these databases cannot be discovered as the current state might depend in some way on the previous states. To this end, a pre-processing of data is needed aimed at extracting relationships intimately connected to the temporal nature of data that will be make available to the discovery algorithm. The predicate logic language of ILP methods together with the recent advances as to efficiency makes them adequate for this task. 1 Introduction Knowledge Discovery in Databases is concerned with identifying interesting patterns in complex structured domains. These databases usually involve several levels of objects and complex relations among them. At the same time new and more efficient ILP algorithms have been developed that are making feasible the application of ILP...

Various knowledge discovery techniques are readily available and many new ones are currently being developed. However, there is no clear understanding of how these techniques compare. This study provides means for the challenging task of analytically comparing knowledge discovery techniques. These means comprise: the notions of applicability, complexity and accuracy of a technique with respect to a problem. We introduce various techniques and problems and derive concrete comparison results. 1 Introduction Knowledge discovery techniques are used to search for relationships and global patterns that exists in large databases but are "hidden" among vast amounts of data. Next generation information systems will incorporate knowledge discovery techniques. Even though various knowledge discovery techniques are readily available and many new techniques are currently being developed, there is no understanding of how these techniques compare. The aim of this study is to define precise notions s...

This paper propose a diagnosis architecture that integrates consistency based diagnosis with induced time series classifiers, trying to combine the advantages of both methods. Consistency based diagnosis allows fault detection and localization without prior knowledge of the device fault modes. Machine learning techniques are able to induce time series classifiers that may be used to identify fault modes of a dynamic systems. The diagnostician performs fault detection and localization resorting to consistency based diagnosis trough possible conflicts. Then, a time series classifier, induced from simulated examples, generates a sequence of faults modes, coherent with the result of the fault localization stage, and ordered by fault modes confidence.

This paper proposes a diagnosis architecture that integrates consistency based diagnosis with induced time series classifiers, trying to combine the advantages of both methods. Consistency based diagnosis allows fault detection and localization without prior knowledge of the device fault modes. Machine learning techniques are able to induce time series classifiers that may be used to identify fault modes of a dynamic systems. The diagnostician performs fault detection and localization resorting to consistency based diagnosis through possible conflicts. Then, a time series classifier, induced from simulated examples, generates a sequence of faults modes, coherent with the result of the fault localization stage, and ordered by fault modes confidence. Finally, to simplify the diagnosis task, it is considered as a subtask of a supervisory system, who is in charge of identifying the working conditions for the physical system.