Modeling Method in the Scientific Research
Faculty of Technology and Metallurgy Belgrade University Karnegijeva, Belgrade, Serbia
To cite this article:
Jelenka Savkovic-Stevanovic. Modeling Method in the Scientific Researc. Science Research. Vol. 3, No. 3, 2015, pp. 66-71. doi: 10.11648/j.sr.20150303.14
Abstract: In this paper modeling method in research was studied. The knowledge acquisition of structural description from examples was provided. Knowledge based systems must represent information abstractly so that it can be stored and manipulated effectively. There are difficulties formulating the knowledge explicitly as rules and other abstractions and induction. Inductive and deductive methods were considered. In this paper models purpose in research work was examined. Knowledge formulation and learning principles by models were studied.
Keywords: Models, Principles, Knowledge Acquisition, Hypothesis, Theory, Testing
Current knowledge acquisition systems perform routine housekeeping, permit rote learning of explicitly presented facts, and are able to elicit from experts simple rules based on the attributes. Methods of concept learning may be able to overcome these imitations, although the present state of the art is primitive and suggests ideas rather than well developed algorithms for the knowledge acquisition tool box -. Concept learning systems take examples and create general descriptions, often expressed as rules, which expert systems need.
Dictionary definitions of concept are remarkably vague, but have in common the abstract idea of a class of objects, particularly one derived from specific instances or occurrences. Learning is an equally broad term, and denotes the gaining of knowledge, skill, and understanding from instruction, experience, or reflection in other words, knowledge acquisition by people. It is taken to denote the acquisition of structural description from examples of what is being described.
Others have defined terms such as generalization inductive learning and inductive modeling in almost identical ways. Moreover, what is learned is sometimes called a generalization, a description, a concept, a model, or a hypothesis. A satisfactory technical vocabulary has not yet been developed, which term one favor seems purely a matter of taste.
Concept learning involves acquiring descriptions that make the structure of generalizations explicit. When a person learns a new concept, he or she gains knowledge that they can use in a rich variety of different ways, apply it to other ideas, and so on. For a person, learning is not simply a matter of acquiring a description, but involves taking something new and integrating it fully with existing though processes. However, they do acquire descriptions that are explicit in the sense that they can be communicated economically for example, in the form of rules, and could plausibly support a variety of different kinds of reasoning. This orientation rules out, for example, connectionist model of learning, which embed knowledge in high dimensional numerically parameterized spaces, making learning a process of weight adjustments -.
While the phrase may conjure up an alluring promise of the magic of human intelligence, systems for concept learning are rooted firmly in the reality of practical algorithms -. The essence is a constrained search of what is invariably an astronomically large space of possible descriptions. The framework can classify into areas of representing concepts and examples, biasing the search for concepts and interacting with a teacher.
In this paper role of model in scientific research was considered.
2. Inductive and Deductive Methods
Two kinds of concept learning are distinguished inductive and deductive. For the inductive concept learning, one must infer a general conclusion from empirical examples . This can be formalized as follows. Along with initial background knowledge -BK, examples -E of a desired concept are given. A concept C is said to be induced if:
BK>E the examples are not logical consequences of the knowledge BK>-C the concept is not inconsistent with the knowledge and examples
BK C> E the examples are logical consequences of the concept and knowledge together
Normally BK>C , C is induced rather than deduced. If E is known to be exhaustive, the learner may be able to deduce a concept. The inference of a statement from information that is known to exhaust all possibilities is a special case of induction known as summative induction, and lies on the borderline with deduction.
Suppose examples and concept are both deducible from background knowledge, even though neither are explicitly, present in it. This is an instance of deductive concept learning. Although, excluded by the above formalization, it can be viewed as a kind of learning, for the examples show which deductions are important namely those that represent the reasoning involved in the examples and enable this important knowledge to operation in an explicit description. In most cases an impossibly large set of operational concepts can be deduced from background knowledge, and the problem is to select one appropriate to the examples. Deductive concept learning may be formalized as follows:
BK=>E the examples are logical consequence of the background knowledge
C BK the concept is not an explicit part of the background knowledge
BK=>C the concept is the logical consequence of the background knowledge
C=>E the examples are the logical consequence of the concept.
In other words, the concept operationalizes the relevant part of background knowledge. This kind of learning is often called explanation based.
The distinction between deductive and inductive concept learning can be viewed as a modern reincarnation of the long philosophical tradition of distinguishing necessary from contingent truths.
Concepts and examples are the output and input of the knowledge acquisition system, what is learned and what is provided by a teacher or other external agent (Fig.1). To be useful, a framework for representing concepts must provide knowledge engineers with methods for selecting appropriate representations for examples, concepts and background knowledge. Separate representations are required for examples and concepts.
Since background knowledge can be represented in the diverse ways, it is hard to categorize its role in biasing search. In practice a different class of methods is used when substantial background knowledge is represented explicitly.
Knowledge sparse techniques are used when little or no domain knowledge is explicitly available during learning. There are three categories: parameter learning, similarity based learning, and hierarchical learning.
Parameter learning optimizes the numerical parameters of a prespecified model. Such system are often designed to detect regularities in noisy situations, and use knowledge in the form of statistical techniques. The structure of the model is not represented explicitly, and the system can not modify it or reason about it.
Similarity based learning delineates the space of possible concept descriptions in advance, searches it for ones that best characterize the structural similarities and or differences between the example presented.
Hierarchical learning augments the description language as each concept is learned. This allows new descriptions to build upon old ones, so that background knowledge grows.
Knowledge rich learning employ considerable knowledge about the target domain. They work by relating new information to the existing body of knowledge. Explanation based learning assimilates a substantial chunk of information by intensive analysis of just one example. Learning by discovery is when systems operate autonomously, performing experiments to enhance their knowledge base.
3. Knowledge Representation
Any one representation will not encompass the broad application of concept learning techniques. There are three techniques: order predicate calculus, expressions composed of functions, and procedures. These reflect fundamental formulations of computing that have been realized in logic, functional and imperative programming styles. Although equivalent in expressive power, the different representations are more or less appropriate for particular concept learning problems, depending on the nature of the examples, background knowledge, the way the complexity of concepts are measured, and the style of interaction with the teacher. For example, decision three are naturally represented as logic expressions, polynomials as functions and robot tasks as procedures. Functional representations incorporate the powerful mathematics available for numbers. Procedures embody the notions of sequencing, side effects and determinism normally required in sequential, real word tasks.
There is an obvious overlap between logical and nonnumerical function representations. For example, the concept of appending lists can equally well be written in logical and functional styles. The difference is that the logical form expression a pure relation without distinguishing input and output, while the functional representation acts on the input list to construct the output.
Many learning methods apply to examples that can be expressed as vectors of attributes in a form equivalent to propositional calculus. This represents logical statements using predicates on constant terms, with connectives for disjunction, conjunctions, negation and implication. The values an attribute can assume may be nominal, linear, or tree structured. A nominal attribute is one whose values form a set with no further structure for examples the set of primary colors. A linear attribute is one whose values are totally ordered for example natural numbers. Ranges of values may be employed in descriptions. A tree structured attribute is one
whose values are ordered hierarchically. Only values
associated with leaf nodes are observable in actual examples: concept descriptions, however, can employ internal node names where necessary.
First order predicate calculus allows variable terms on logical statement and quantification over those variables. Attribute vectors, propositional calculus are not powerful enough to describe situations where each example comprises a scene containing several objects. Objects are characterized by their attributes. Moreover, pair wise relations may exist between them. This means that variables must be introduced to stand for objects in various relations. Such relations can be described by predicates which, like attributes, may be normal linear or tree structured. Objects and concepts are characterized by combinations of predicates.
Functional expressions include many natural laws, as well as relationships between quantities and parameters. Functional representations are appropriate for nested and recursive numeric or nonnumeric expressions. Any functional relationship can be represented in logic and this is no surprise since the two forms are expressively equivalent. But in a framework for induction, it is preferable to treat functional expressions separately and omit explicit quantifiers. An important difference is that functional representations of concepts must be single valued, while logical expressions do not need to be this greatly affects the search space involved.
A more suitable form of representation might be the functional calculus, or some incarnation of it in pure of it in functional programming languages. Work on programming language semantics, also partly based on functional calculus and often coupled with function programming languages may suggest appropriate forms of expressly prohibit aspects that distinguish functional from procedural representations in the framework, namely side effects and reliance in sequential execution. Work on programming language semantics, also partly based on functional calculus and often coupled with function programming languages may suggest appropriate forms of representation. Existing function induction systems are specially designed for particular domains with little attention to more general forms.
Typical concepts in procedures category include procedures for assembly welding, and standard office procedures. The procedural formalism suitable for representing sequential execution where side effects, such as variable assignment and real world outputs like movements, make it vital to execute the procedure in the correct order. To describe a procedural language formally, a binding environment model of execution is needed, instead of the simpler substitution model that suffices for pure functional representations must be deterministic to be useful procedural concepts.
Note the generalization of execution traces into a procedure is not reducible to an equivalent problem involving the generalization of nonsequential input/output pairs as such a reduction will lose information about sequential changes in state.
4. Real System Modeling
A cybernetic approach in model building is shown in Fig.2.
The ost important distinction in a knowledge acquisition framework is how systems represent what they learn. Knowledge representation has always been a central topic in learning process . Presentation schemes identify logic procedural representation, semantic networks, production systems, direct analogical representations, semantic primitives, and frames and scripts. It is hardly necessary to say that a data structure is not knowledge, any more than an encyclopedia is.
For both inductive and deductive concept learning the formalism above suggests representing concepts in an appropriately powerful form of logic, such as predicate calculus. However, although formal logic provides a sufficient basis for deduction, as a foundation for induction it is at once too narrow and too powerful. Extending the distinction that is often made between the epistemological adequacy of a representation, whether it is capable of expressing the facts that one has about the world and its heuristic adequacy, one might characterize formal logic as epistemologically adequate but inductively inadequate for concept learning. On the other hand, logic is too powerful because the need to acquire knowledge automatically from teacher or environment and integrate it with what is already knows means that only the simplest representations are used by programs for system learning.
A model for conflict resolution algorithmic based is shown in Fig. 4.
5. Background Knowledge
How much knowledge does a system need before it can infer concepts from examples? There is an often quoted aphorism that on order to learn something you must nearly know it already. The amount of knowledge already possessed about a problem domain critically affects the kind of things one can expect to be learned. Concept learning techniques can be classified as knowledge sparse and knowledge rich, not so much on the basis of the amount of knowledge they embody as on the extent to which that knowledge is represented explicitly. Even the first kind embodies substantial prior knowledge in the form languages in which examples and concepts are expressed. It is obviously desirable
for a knowledge acquisition system to learn most of
what it needs to know, otherwise its utility will be outweighed by the burden of expressing the prior knowledge. Unfortunately existing concept learning systems are able to learn very little, if any of the background knowledge needed for them to work.
Area concerns the cooperation and pedagogical skills of the teacher is very important. It is not possible, to supply a comprehensive set rules to determine which concept learning scheme to use, given desire concept and example representations, background knowledge available, form of teacher interaction, and appropriate biases for concept representations. This paper leads suitable choices by providing examples of particular kinds of system.
Faced with practical problem, the first decisions to make are how to represent concepts and examples. Suitable forms of concepts representation will be dictated by the requirements of the process based system and the kind of examples available. Sometimes the example representation dominate the decision, while in other situations the desired concept format will force examples into a particular mode.
Logical representations are indicated by the predominance of logical relationships in example or concepts. The possibility of using attribute values strongly suggests the simpler prepositional calculus representation. Similarity based system methods are appropriate when many examples are available, or when it is not possible to define a domain theory in advance. If concept must be built on earlier learned ones, a hierarchical method is indicated. If a domain theory is known, then an explanation or discovery based system can utilize it.
Given a set of objects that represent examples and counterexamples of concept, a similarity based learner attempts to induce generalized description that encompasses all the examples and none of the counterexamples. Interesting general issues include the conditions under which the procedure converges to a single description, weather the system can know that it has converges, whether the final concept may depend on the order of presentation of examples, and whether the training sets expected to be exhaustive or representative.
The version space approach to concept learning transforms the inductive problem of generalization into a deductive one by circumscribing the way in which descriptions are expressed and searching for ones that fit the examples given. It postulates a language in which objects are expressed. Given asset of positive and negative examples of a target description, a simple search algorithm exists that finds all descriptions that are consistent with the examples (Fig.1). This set is called the version space.
The method applies simply and directly when each object is described by a fixed set of properties, usually represented as an attribute vector, which is equivalent to a description in propositional logic. Its performance in such domains has been studied extensively. Allowing disjunction in the description language causes the version space to explode, while even with purely conjunctive concepts, one version space boundary can grow exponentially in the number of examples.
In this paper modeling in scientific research was studied. Model of the scientific research work was derived. Conflict resolution models building was considered.
Knowledge representation and process learning in the scientific research were examined. It provides a general bases for thinking about concept learning in knowledge acquisition problems, techniques and applications. Different methods for truth seeking out were studied and compared.
The author wishes to express her gratitude to the Fund of Serbia for financial support.