Machine Learning Methods for Commonsense Reasoning Processes: Interactive Models

Machine Learning Methods for Commonsense Reasoning Processes: Interactive Models

Xenia Naidenova (Military Medical Academy, Russia)
Indexed In: SCOPUS
Release Date: October, 2009|Copyright: © 2010 |Pages: 424
ISBN13: 9781605668109|ISBN10: 1605668109|EISBN13: 9781605668116|DOI: 10.4018/978-1-60566-810-9

Description

The reduction of machine learning algorithms to commonsense reasoning processes is now possible due to the reformulation of machine learning problems as searching the best approximation of a given classification on a given set of examples.

Machine Learning Methods for Commonsense Reasoning Processes: Interactive Models provides a unique view on classification as a key to human commonsense reasoning and transforms traditional considerations of data and knowledge communications. Containing leading research evolved from international investigations, this book presents an effective classification of logical rules used in the modeling of commonsense reasoning.

Topics Covered

The many academic areas covered in this publication include, but are not limited to:

  • Artificial Intelligence
  • Commonsense reasoning
  • Coordination of commonsense reasoning operations
  • Deductive-inductive commonsense reasoning
  • Expert system generation
  • Human commonsense reasoning processes
  • Knowledge in the psychology of thinking
  • Machine Learning
  • Modeling conceptual reasoning
  • Object-oriented technology
  • Psycho-diagnostic systems generation
  • Reasoning in intelligent computer systems

Reviews and Testimonials

This book demonstrates the possibility of transforming a large class of machine learning algorithms into integrated commonsense reasoning processes in which inductive and deductive inferences are not separated one from another but moreover they are correlated and support one another.

– Xenia Naidenova, Military Medical Academy, Russia

Table of Contents and List of Contributors

Search this Book:
Reset

Preface

TENTATIVE

The main purpose of this book is to demonstrate the possibility of transforming a large class of machine learning algorithms into integrated commonsense reasoning processes in which inductive and deductive inferences are not separated one from another but moreover they are correlated and support one another. This class of learning algorithms embraces the inferring of functional, approximate and implicative dependencies, rough sets, the inferring of concepts and decision trees from a given set of examples and many others knowledge acquisition tasks related to extracting from data and using hierarchical classifications (ontologies) and logical dependencies between them. Consequently, we consider extracting and using conceptual knowledge the main elements of which are objects, properties (attribute values) and classifications (attributes). Taking into account that implications express the links between concepts (object ¡ê class, object ¡ê property, property ¡ê class) we believe commonsense reasoning to be based on using and searching only one type of logical dependencies, namely, implicative dependencies.

The reduction of machine learning algorithms to commonsense reasoning processes became possible due to the reformulation of machine learning problems as searching the best approximation of a given classification (a partitioning) on a given set of examples. The concept of a good classification (diagnostic) test underpins our approach to modeling commonsense reasoning processes combining deductive an inductive inferences. Classification test has a dual nature: from one side it is a logical expression in the form of implication or functional dependency (strict or approximate one), from the other side, it generates a partitioning of data (observations or objects) into disjoint classes.

The theory of algebraic lattices is used as the mathematical language for constructing algorithms for inferring all kinds of good classification tests. The advantage of algebraic lattice consists in the fact that it is defined both as a declarative structure of knowledge representation and as a system of dual lattice operations with the use of which the lattice elements are generated. The processes of inferring chains of lattice elements ordered by inclusion relation serve as a basis for generating all kinds of classification tests. Four variants of inductive transitions from a lattice element to its nearest one in a chain are determined. The special reasoning rules for realizing these transitions are introduced: generalization, specialization, diagnostic rules, and dual diagnostic rule.

In order to show that these rules of inductive transitions are reduced to commonsense reasoning rules, it was necessary to consider commonsense reasoning (deductive and inductive) as a system of logical rules of the two types. The rules of the first type are implicative logical assertions (classification tests) with the use of which the dependencies between observed objects or situations are described. The rules of the second type are reasoning rules (deductive and inductive ones) by means of which the rules of the first type are used, modified or inferred from observations (data) and the rules of the first type (knowledge) .

Deductive rules of the second type are based on the reasoning rules using implications. These rules are modus ponens, modus ponendo tollens, modus tollendo ponens, and modus tollens. Inductive rules of the second type are the known inductive methods of J. S.Mill: the Method of Agreement, the Method of Difference, and the Joint Method of Agreement and Difference. The analysis of lattice construction algorithms allowed to reveal that the rules of inductive transitions are reduced to the commonsense reasoning rules of the second type. The rules of the first type are generated during lattice construction and they are involved immediately via deduction in the inference of classification tests in order to reduce the space of search for new implicative rules (tests) of the first type.

We use also the decomposition of inferring good classification tests into subtasks that allow realizing incrementally commonsense reasoning processes.

Process of commonsense reasoning is considered as a sequence of the following mental acts: implementing a reasoning rule (deductive or inductive), tuning the boundaries of the search space, and choosing a new reasoning rule or subtask. Our model of commonsense reasoning implies a new approaches to many connected problems of data and knowledge integration in intelligent computer systems. These problems are considered in this book.

The first chapter gives a view of historical developing the concepts of knowledge and human reasoning both in mathematics and psychology. The mathematicians create the formal theory of correct inferences, the psychologists study cognitive mechanisms that underpin knowledge construction and thinking as the most important functions of human existence. They study the actual workings of human minds. The progress in understanding human knowledge and thinking will be in combining the efforts of scientists in these different disciplines. We treat knowledge and reasoning not as independently defined problems and we strive, in this chapter, to cover the central ideas of knowledge and logical inference that have been manifested in the works of outstanding thinkers and scientists of the past time.

In the second chapter we focus on the tasks of knowledge engineering related mainly to knowledge acquisition and modeling integrated logic-based inference. We have overlooked the principal and more important directions of researches that pave the ways to understanding and modeling human plausible (commonsense) reasoning in computers.

The third chapter develops a conception of commonsense reasoning based on mutually coordinated operations on objects, classes of objects, and properties of objects. This conception goes back to the model of classification processes given by J. Piaget, the outstanding psychologist of the XX century.

The operations of classification are the integral part of any reasoning about the time, the space, the things, the events, the motion and so on. They consolidate all forms of reasoning and they make it possible to present knowledge as the system of interconnected relations.

In forth chapter we describe a model of commonsense reasoning that has been acquired from our numerous investigations on the human reasoning modes used by experts for solving diagnostic problems in diverse areas such as pattern recognition of natural objects (rocks, ore deposits, types of trees, types of clouds e.t.c.), analysis of multi-spectral information, image processing, interpretation of psychological testing data, medicine diagnosis and so on. The principal aspects of this model coincide with the rule-based inference mechanism that has been embodied in many expert systems.

The fifth chapter contains some examples of natural human commonsense reasoning both in scientific pattern recognition problems and in solving logical games. An analysis of inference structure shows that inductive and deductive rules communicate in reasoning. An automated model for detecting the types of woodland from incomplete descriptions of some evidences is also given in this chapter. An interesting part of this model is a small knowledge base based on the representation of experts¡¯ knowledge about natural woodlands as biological formation.

The sixth chapter discusses a revised definition of classification (diagnostic) test. This definition allows considering the problem of inferring classification tests as the task of searching for the best approximations of a given classification on a given set of data. Machine Learning methods are reduced to this task. An algebraic model of diagnostic task is brought forward founded upon the partition lattice in which object, class, attribute, value of attributes take their interpretations

In the seventh chapter, the definition of good diagnostic test and the characterization of good tests are discussed and the concepts of good maximally redundant tests (GMRTs) and good irredundant tests (GIRTs) are introduced.

The definition of the good test is based on the partition model of classifications that has been given in the previous chapter. Some characteristics of good tests are determined a strategy for inferring all kinds of good diagnostic tests. We describe an algorithm called Background Algorithm based on the method of mathematical induction. This algorithm is applicable to inferring all kinds of good classification tests and, consequently, for inferring functional, implicative dependencies and association rules from a given data set. We discuss also, in this chapter, the possible ways of constructing an efficient algorithm for inferring good tests of any kind.

The concept of good classification test is redefined in the eighth chapter as an element of a dual lattice. The operations of lattice generation take their interpretations in human mental acts. Inferring the chains of dual lattice elements ordered by the inclusion relation lies in the foundation of generating good classification tests. The concept of an inductive transition from one element of a chain to its nearest element in the lattice is determined. The special reasoning rules for realizing inductive transitions are formed. The concepts of admissible and essential values (objects) are introduced. Searching for admissible or essential values (objects) as a part of reasoning is based on the inductive diagnostic rules. In this chapter, we also propose a non-incremental learning algorithm NIAGaRa based on a reasoning process realizing one of the ways of dual lattice generation. Next we discuss the relations between the good test construction and the Formal Concept Analysis (FCA).

The most important steps in the direction to an integrative model of deductive-inductive commonsense reasoning are made in the ninth chapter. The decomposition of inferring good classification tests is advanced into two kinds of subtasks that are in accordance with human mental acts. This decomposition allows modeling incremental inductive-deductive inferences. We give two basic recursive procedure based on two kinds of subtasks for inferring all good maximal redundant classification tests (GMRTs): ASTRA and DIAGaRa. An incremental algorithm INGOMAR for inferring all GMRTs is presented too. The problems of creating an integrative inductive-deductive model of commonsense reasoning are discussed in the last section of this chapter.

The tenth chapter summarized some methods of inferring approximate diagnostic tests. Approximate minimal diagnostic tests allow constructing a model of commonsense reasoning by analogy. The system ¡°DEFINE¡± for analogical inference and some results of its application are described. Mining approximate functional, implicative dependencies and association rules is based on the same criteria and on the application of one and the same algorithm realized in the Diagnostic Test Machine described shortly in this chapter. Some results of inferring ¡°crisp¡± and approximate tests with the use of Diagnostic Test Machine are give in Appendix to this chapter. The eleventh chapter deals with the description of the possible mechanisms for data-knowledge organization and management in intelligent computer systems.

We will see intelligent computer system as a system capable to commonsense reasoning. The database system will be intelligent if it communicates with users via conceptual knowledge and by the use of commonsense reasoning on knowledge but not via special formal queries to data. In this chapter, we attempt to consider the main principles which could be posed in the foundation of constructing intelligent computer database systems first of all.

A technology for rapid prototyping and developing expert systems or intelligence systems as a whole is proposed in the twelfth chapter. The main parts of the technology are the object-oriented model for data and knowledge representation and the mechanism for data-knowledge transformation on the basis of an effective algorithm of inferring all good classification tests. An approach to expert system development by means of the technology proposed is analyzed. The tool-kits for expert system generation are described and the application of these tools is demonstrated for development of a small geological expert system.

The thirteenth chapter proposes an automated technology for creating the applied psycho-diagnostic expert systems (APDS) the main peculiarity of which consists in using the machine learning methods to choose, validate, define and redefine the main constructive elements of testing and problem-solving methods used in psycho-diagnostic systems.

Author(s)/Editor(s) Biography

Xenia Naidenova is a senior researcher of the Group of Psycho Diagnostic Systems’ Automation at the Military Medical Academy (St. Petersburg, Russia). She is currently the head of Project DIALOG: Methods of Data Mining in Psychological and Physiological Diagnostics. Dr. Naidenova received a diploma of engineering with a specialty in computer engineering (1963) and a PhD in technical sciences (1979), both from the Lenin Electro-Technical Institute of Leningrad. In 1999 she received a senior researcher diploma from the Military Medical Academy (St. Petersburg, Russia). She has guided the development of several program systems on knowledge acquisition and machine learning including DEFINE, SIZIF, CLAST, LAD, and diagnostic test machines and has published over 150 papers. Dr. Naidenova is a member of the Russian Association for Artificial Intelligence and is on the Program Committee for the KDS.

Indices