Improving Knowledge Based System Performances by Experience

The proposed approach presents several improvements over the existing ones. In particular, it allows the reasons of the occurred failures to be explicitly analyzed, possibly taking into account a-priori information. Moreover, the representation of the knowledge can be quite complex: the concept descriptions are expressed in a language derived from First Order Predicate Logic, augmented with numerical quantifiers. The rules used to discriminate concept instances are grouped into clusters, organized in a graph structure and each rule is affected by a reliability measure.

The learning methodology is knowledge intensive, and makes use of a set of updating procedures, to be applied under the supervision of a reasoning system provided with its own knowledge. As the system has been designed to work in real domains, usually affected by noise, uncertainty and errors, particular attention has been given to the computational complexity problem. The system has been tested on a problem taken from the field of speech recognition.

Constructive Learning in Conjunctive Concept Characterization

The goal of the present paper is that of analyzing the impact of a special kind of constructive learning to the problem of generating the Maximally Specific Characterization of a set of structured examples; this kind of constructive learning concerns the introduction of numerical quantification. A set of algorithms, which achieve the goal, is also presented. The design of these algorithms was biased by the requirement of low computational complexity. The needs of computational resources is here traded-off with the level of details (i.e. the descriptive power of the representation language) the characterization is supposed to supply: the more detailed the characterization, the higher the needed amount of resources.

The presented algorithms perform a one-step learning of a concept characterization, starting from a set of positive instances and utilizes the simultaneous availability of these instances to reduce the search space. Then, given a set F of examples, the obtained characterization does not depend on any ordering among them. An example of application of the methodology is also presented.

Handling the Extensional Representation of Fuzzy Logic Formulas

Knowledge Base Refinement Using a Domain Theory

The novelty of the method consists in the use of a body of deep knowledge for guiding the process of rule refinement, even in case this deep knowledge is too complex or not specific enough to directly generate useful classification rules (as it is done in Explanation Based Learning). In this case, it will be shown that such kind of theory can still be used for finding partial justifications for the currently available classification rules.

Justifications can be exploited in order to localize failures and to propose changes, in agreement with the theory, oriented to improve the knowledge base. If only incomplete justifications are found, statistical evidence is used to select those which are likely to be the most reliable ones; justifications are also used to evaluate proposed changes in the knowledge base; this is done by reasoning on many examples and counterexamples at the same time and an algorithm for performing the knowledge base refinement is also presented.

Knowledge Refinement in Diagnostic Expert Systems

Justifications derived from the deep knowledge can be exploited in order to localize failures and to propose changes oriented to improve the knowledge base. If only incomplete justifications are found, statistical evidence is used to select those which are likely to be the most reliable ones; justifications are also used to evaluate proposed changes in the knowledge base; this is done by reasoning on many examples and counterexamples at the same time.

Automated Knowledge Acquisition in Expert Systems

Improving Learning Using Causality and Abduction

Use of Causal Model to Learn Diagnostic Knowledge in a Real Domain

Learning Quantitative Features in a Symbolic Environment

This paper present a method for dealing effectively with numeric features in a learning framework based on a first order logic. The method has been implemented and works in two steps. In the first step, tentative numeric features are learned together with symbolic features, whereas in the second one the numeric knowledge is refined using a standard genetic algorithm.

The implementation is evaluated on a problem of pattern recognition, already described in the literature by the authors. The results are encouraging and show the viability of the approach.

Improving Learning by Using Deep Models

The problems of imperfection and intractability of the theory are handled by allowing the system to make assumptions during its reasoning. In this way, robust knowledge can be learned with limited complexity and limited number of examples. The system has been applied to learn the knowledge base of a diagnostic expert system for mechanical trouble-shooting, allowing comparisons with a previous expert system for the same task, whose knowledge base has been automatically acquired without a deep model, to be performed.

An interesting aspect of the system is the tentative identification of the performance element with the learner, through the common use of the deep knowledge, of the learned knowledge base and of a justification structure which the system builds up and maintains.

Use of Causal Models and Abduction in Learning Diagnostic Knowledge

Learning in Uncertain Environments

Learning Behavioral Knowledge in Robotics Domains

The problems of imperfection and intractability of the theory are handled by allowing the system to make assumptions during its reasoning. In this way, robust knowledge can be learned with limited complexity and limited number of examples. For the sake of illustration, an simple example where a robot is asked to learn whether an object is dangerous because hot, or not, is presented.

Comparison of Search Strategies in Learning Relations

The results show that the information gain criterion is too greedy to be useful when the concepts have a complex internal structure; however, this drawback is more or less shared with any purely statistical evaluation criterion. The addition of parts of the available domain theory increases the obtained performance level.

Learning Relations: An Evaluation of Search Strategies

The results show that the information gain criterion is too greedy to be useful when the concepts have a complex internal structure; however, this drawback is more or less shared with any purely statistical evaluation criterion. The addition of parts of the available domain theory increases the obtained performance level. Similar results have been previously obtained on a number of real applications and of test-cases taken from standard machine learning data bases.

Multi-Strategy Learning and Theory Revision

Learning Fuzzy Concept Definitions

Smart+: A MultiStrategy Learning Tool

WHY: a System That Learns Using Causal Models and Examples

Learning First Order Theories

Moreover, the learning system described in the paper can be used both to learn new knowledge from scratch and to refine an existing background theory.