Transcript Prezentacja programu PowerPoint

Leszek Pacholski
ENGINEERING WITH COMMERCE
Intelligent Management
Support Systems
1. Choice of methodology of the management intelligent support in
regard to formal complexity of the phenomenon and stage of
understanding the dynamics of its course.
2. Artificial Inteligence – differences between traditional software
and A.I. software.
3. Structure of an Expert System.
4. Instruments for creation of Expert System.
5. Informal methods of representation of knowledge.
6. Classical methods of managerial reasoning used in Knowledge
Bases.
7. The concept of Fuzzy Sets in processes of managerial reasoning.
8. Course of Fuzzy Reasoning.
9. Artificial Neural Networks – definition, directions of applications.
10. The Kohonen Maps: Self Organizing Maps & Learning Vector
Quantization.
11. The Recurrent Networks – exemplary architectures.
12. The Algorithms of Evolutional Processing.
1.Choice of methodology of the management intelligent
support in regard to formal complexity of the phenomenon
and stage of understanding the dynamics of its course.
COGNISION
.
latent overt -
precize -
Inductive reasoning
Algorythmic
reasoning
Precize understanding
of problems dynamics
Deductive
reasoning
Non-observable constant
paths of behaviour
Observable constant
paths of behaviour
Expert systems
Parapetric regression
Non-parametric regression
Artificial neural networs
Adaptive fuzzy systems
Genetic algorythms
COMPLEXITY
Algorythmic reasoning is used in cases in which there is a possibility of precise
understanding and describing the phenomenon (there is knowledge enabling to
create a model of particular phenomenon in a form of clearly defined and
determined dependencies),
- Deductive reasoning can be applied to phenomenons that we can’t precisely
describe in a mathematical form, yet basing direct observation we can find
some stable patterns of determined phenomenon,
- Inductive reasoning is used when there is no possibility of direct precize
determining constant patterns of the course of particular phenomenon.
-
2a. Artificial Intelligence. Differences between traditional
software and artificial intellgence software
Artificial Intelligence is a domain of computing science
relating to methods and techniques of symbolic reasoning
realized by a computer with help of symbolic
representation of knowledge used
in the process of reasoning.
The essence of artificial intelligence software is
ascertainment of indequacy of most of modern computer
systems with phenomenons
apearing in real environment.
Mentioned inadequacy concerns using two-value logic for
description of so-called fuzzy phenomenons. Such practice
leads to limitation of areas of efficient support and
increases costs of introducing computers in practice.
2b. Differences between traditional software and A.I. software.
Traditional software programs
A. I. software programs
Digital processing
Symbolic processing
Algorythmic record of operations
Declarative record of knowledge
Batch processing or interactive processing
Interactive environment of integrated
development environment
Capability of verification of correctness
Lack capability of full verification of
of operation of software
correctness of operation of software
Development of software on base
Development of software on basa of
of specification
creation of prototypes and improving
them
Presentation and use of data
Presentation and use of knowledge
Utilization of databases
Utilization of knowledge
3a. Structure of an Expert System
Expert of particular
domain
Knowledge engineer
Interface
Database
Subsystem
collecting
knowledge
Knowledge base
Expert system
Explaining
subsystem
Concluding
subsystem
Interface
User
3b. Structure of an Expert System
Knowledge engineer gathers knowledge from experts, transforms it
acoordingly to methods of its prezentation in knowledge base and
puts knowledge processed in such way into the knowledge base.
Knowledge base stores primal knowledge, procedural rules and
experiences necessary for expert system functioning.
Concluding subsystem enables creating new knowledge based on
existing one.
It is based on following features in its research of knowledge:
- progressive reasoning (aiming to the required target from known
conditions),
- regressive reasoning (aiming to necessary targets from required
conditions).
Explaining subsystem presents operations of an expert system
through a list of rules used while the expertise was being formed.
4. Instruments for creation of expert system
Expert system tools can be divided into:
1. Expert system shells
2. Environmental programs facilitating implementation
system (ex. programs facilitating knowledge base
management, facilitating management of the
knowledge base, graphic editors)
3. Expert system languages (Clips, Flops, OPS5)
4. Symbolic programing languages (Lisp, Prolog)
5. Algorythmic languages (Basic. Pascal, C, C++ ...)
5a. Informal methods of representation of knowledge - conclusions
Conclusions are one of most important elements of knowledge base.
They concern: events, phenomenons, symptoms, activities.
Usuallu they are recorded in form of an ordered tripple:
(<OBJECT>, <ATTRIBUTE>, <VALUE>)
Dictionaries of names of objects and attributes and their values are
used for simplification of proposition forms.
In some systems conclusions are presented in form of ordered four :
(<OBJECT>, <ATTRIBUTE>, <VALUE>,<CF>)
where CF is a Cerntainty Factor leading to a so-called approximate
conclusions.
Levels of certainty are determined subjectively (ex. in an interval:
[0,1] or [-1,1])
5b. Informal methods of representation of knowledge Vectors of Knowledge & Semantic Nets
Vectors of knowledge are some kind of generalization of rule
methods of presentation of knowledge, for which rules are being
recorded in form of vectors.
The method uses three symbols illustrating knowledge:
* – particular condition/conclusion doesn’t
appear in the rule at all,
T – particular condition/conclusion is true,
N – particular condition/conclusion is fals.
Semantic Net is a sort of logic, which presents in a graphic form
relations between objects within this logic.
In other words it is a kind of drawing of a deduction mechanism.
Concluding corresponds here with „moving” on the illustration.
Various conclusions result from an inspection of the net.
5c. Informal methods of representation of knowledge - Frame Systems
The idea of frame systems is based on analysis of man’s behaviour
in situations he has never experienced and in new environment, of
which he was having already some concept.
In such situations man taking out of his memory particular
structure, i.e. frame, than he confronts the situation with the
knowledge from the frame.
Yet, when man meets a totally new object, his first reaction would
be a trial of remembering and naming this new phenomenon (i.e.
creating a new frame).
6. Classical methods of managerial reasoning used in knowledge bases.
The idea of PROGRESSIVE REASONING consists generation of
new facts (on base of accessible rules and facts) as long as the put
aim (for example a hypothesis) would appear in the group of
generated facts.
The basic feature of such method of reasoning is the ability of
increasing the base of facts.
REGRESSIVE REASONING reasoning is about presenting
veracity of the main hypothesis basing on authenticity of
prerequisites.
If we don’t know if the prerequisite is true, we treat it as a new
hypothesis and we try to prove it.
In resultlt of such operations we would find a rule, in which all
prerequisites are true, the conclusion of such rule would be also true.
Such conclusion would be a basis for next rule, which has an
unknown earlier prerequisite etc.
Presented hypothesis is true if all considered prerequisites are
presentable.
7. The concept of Fuzzy Sets in processes of managerial reasoning.
The definition of fuzzy set is generalizing the term of classic set, i.e.
allowing the determing function (so-called membership function)
to obtain values of extremal states of determined set
(one or zero {0,1}), as well as intermediate values from this range
(interval [0,1]).
So, in a fuzzy set we observe a fluent passage from absolute
membershipto no membership at all.
Elements can also be part of the set in a certain degree.
The concept of fuzzy sets is often used in processes of managerial
reasoning.
This idea, in the range of representation of knowledge, describes the
problem with help of fuzzy sets theory because tresspassing limits
of Aristotle’s logic rules enables better modeling of decisiv limits for
fuzzy verbal terms.
8. Course of fuzzy reasoning
Determine values of function of membership
for individual fuzzy notions appearing in conditions of rules
Determine fuzzy areas referring to variable included
in rule’s conclusion basing on function of membership
Make the combination of fuzzy areas determined
during former step
Basing on obtained fuzzy areas determine the result fuzzy area
Make the defuzzyfication (change the fuzzy set into a certain
numerical value) of the result fuzzy area.
9a. Artificial Neural Networks – definition, directions of applications
Artificial Neural Networks remain a system of simple elements
processing information linked together, called neurons, units or nods.
ANN are classified as learning systems.
There are weight factors assigned with each connection between
elements, which determine the volume of those connections and
define the set of parameters of this model.
Weigh factors are assigned or determined by a training process,
which aims to teach ANN to identify patterns or project transitions.
Neural network obtains a certain structure. Its units are grouped into
so-called layers (ANN architecture). Neural network is characterized
by: - network’s architecture, i.e. localization of individual neurons
and connections between them, - the process of searching, i.e.
the method of transfer of information (from input to output),
- the method of learning (training), which determine the initial set of
weigh factors and the way, in which those factors should be changed
along the process of learning.
9b. Artificial Neural Networks – definition, directions of applications
Unlike classical computing systems, ANN operate as learning
systems and it is possible to separate two stages: training stage and
stage of reaction to particular external stimulus.
At first the model of solution might be unknown, hence it should be
build by the network in its process of learning, basing on so-called
training information that it has obtained. Such approach causes
many changes in way of designing and building ANN systems,
in comparison to traditional computing systems.
In an approach typical for ANN calculations are realized
by processing units-neurons.
Each neuron is connected with a certain number of other neurons
and becomes a part of a dispersional system.
The ANN domain is also called connectionism, because it keeps
information inside a structure of connections between neurons.
Artificial neural network is not being programmed.
It is being thought.
9c. Artificial Neural Networks – definition, directions of applications
Most often artificial neural networks realize following
sorts of processing:
- reminding (restoring or interpreting) of information
stored in ANN system,
- classification,
- recognizing,
- estimation (approximation, interpolation, filtration,
forecasting, predicting),
- optimization (including solutions of linear and
non-linear equations),
- intelligent steering (without the necessity of preparing
a model, based exclusively on experience).
10a. The Kohonen Maps: Self Organizing Maps & Learning Vector
Quantization
The Kohonen Maps are a specific group of neural
networks, which usually has a function of classifiers that
relatively simplify data grouping.
They consist of an input layer and one layer of processing
neurons.
Each neuron in Kohonen layer is connected with all inputs.
Kohonen Maps use an algorithm of learning, also called
competitive learning.
Neurons in the network compete with each other in
response to the input signal.
The winning neuron and its closest environment learn in a
process that rest on approximation of theirs weights to the
input vector.
10b. The Kohonen Maps: Self Organizing Maps & Learning Vector
Quantization
The most commonly used type of Kohonen maps are so-called SOM
(Self Organizing Maps), i.e. network maps. Competitive layer in
those networks has usually a form of two or three dimensions table
of neurons. Neurons are connected with all inputs, so each neuron
has as many weight factors as number of inputs the system consists.
10c. The Kohonen Maps: Self Organizing Maps & Learning Vector
Quantization.
The most commonly used type of Kohonen maps are so-called LVQ
networks (Learning Vector Quantization).
They illustrate a controlled competitive learning.
Each training vector must be labeled by a tag of its class. Particular
tags in LVQ network are assigned to one or several neurons in its
competitive layer. Each unit is connected with all inputs.
11a. The recurrent networks - exemplary architectures.
TheThe topology of recurrent networks allows using
regressive connections.
The input signal from any unit can be transmitted to its
input (directly or with help of other neurons).
So the state of neuron becomes dependent not only from
the value of the input signal, but also from a state of any
unit, including the neuron itself.
The most simple example to present is a modification of
one direction network from the type of regressive
propagation through adding to the input layer units of socalled context.
11b. The recurrent networks - exemplary architectures.
12. The algorithms of evolutional processing.
Models of algorithms of evolutional processing solve optimizational
problems and tasks of searching with help of a method similar to the
rule of real evolutional mechanism, i.e. Darwin’s strategy of survival
of most adapted individuals.
Such algorithms might be used in practical tasks for difficult
questions of optimization and for research in cases when
numeric or heuristic solutions are not possible to obtain or if those
solutions drive to unsatisfactory results.
Three basic streams of evolutional algorithms concern as follows:
genetic algorithms, evolutional strategies, evolutional programming.
Strategy of an evolutional algorithm consists translocating from one
population of solutions into another one, next to the preceding
population, in accordance to certain rules that have been established.
An evolutional algorithm realizes generating next population
and searching it and selecting new generation through following
genetic operators: selection, crossing or mutation.