Upload
oded
View
36
Download
0
Embed Size (px)
DESCRIPTION
MITM 613 Intelligent System. Chapter 9: Hybrid Systems. Chapter Nine: Hybrid Systems. 9.1 Convergence of techniques 9.3 Genetic-fuzzy systems 9.4 Neuro-fuzzy systems 9.5 Genetic-neural systems. Many computing techniques can be applied to particular problems: - PowerPoint PPT Presentation
Citation preview
Abdul Rahim Ahmad
MITM 613Intelligent System
Chapter 9: Hybrid Systems
Chapter Nine: Hybrid Systems
Abdul Rahim Ahmad
2
9.1 Convergence of techniques9.3 Genetic-fuzzy systems9.4 Neuro-fuzzy systems9.5 Genetic-neural systems
Convergence of techniques
Many computing techniques can be applied to particular problems:
symbolic representations (eg. knowledge-based systems)
computational intelligence
conventional programs
They are not exclusive alternatives but complementary to each other and can be made into hybrid of mixed techniques.
Abdul Rahim Ahmad
3
Convergence of techniques
How different computational techniques can be complementary? When dealing with multifaceted problems
For Capability enhancement
For Cases where parameter settings are needed
For Clarification and verification
Abdul Rahim Ahmad
4
Combining Techniques
When Dealing with multifaceted problems Real life problems are multi facet.
Different specialize module deal with specific task
Modules can communicate between each other (eg: in Black board system)
When Enhancing Capability of a system A technique A can be used within another
technique B to enhance the capabilities of B.
Abdul Rahim Ahmad
5
Example :hybrid as a blackboard system
Abdul Rahim Ahmad
6
Combining Techniques
When Setting Parameters Use one technique to set the parameters of another
technique.
Eg: neuro-fuzzy, genetic-fuzzy, genetic-neural.
When Performing Clarification and verification (In Neural network)
Hybrid method used to find the reasons of associations bet. input and output and the weights.
To automatically extract from the network, rules that can be readily understood, and to write verification rules to check the validity of the network’s output.
Abdul Rahim Ahmad
7
9.2 Blackboard System
Abdul Rahim Ahmad
8
Blackboard Model
Blackboard model or blackboard architecture provides a software structure well suited to multifaceted tasks.
Systems that have this kind of structure are called blackboard systems. In such a system, knowledge of the application domain is divided into
modules, referred to as knowledge sources (or KSs), each of which
is designed to tackle a particular subtask. KSs are independent and can communicate only by reading from or
writing to the blackboard, a globally accessible working memory where the current state of understanding is represented.
KSs can also delete unwanted information from the blackboard.
Abdul Rahim Ahmad
9
Blackboard System
A blackboard system is analogous to a team of experts who communicate their ideas via a physical blackboard, by adding or deleting items in response to the information that they find there.
Each knowledge source represents such an expert having a specialized area of knowledge.
As each KS can be encoded in the most suitable form for its particular task, blackboard systems offer a mechanism for the collaborative use of different computational techniques such as rules, neural networks, and fuzzy logic.
The inherent modularity of a blackboard system is also helpful for maintenance. Each rule-based knowledge source can use a suitable reasoning strategy for its particular task, e.g., backward- or forward-chaining, and can be thought of as a rule-based system in microcosm.
Abdul Rahim Ahmad
10
Knowledge sources are applied in response to information on the blackboard, when they have some contribution to make. This leads to increased efficiency since the detailed knowledge within a knowledge source is only applied when that knowledge source becomes relevant.
In the idealized blackboard system, the KSs are said to be opportunistic, activating themselves whenever they can contribute to the global solution.
However, this is difficult to achieve in practice as it may involve interrupting another knowledge source that is currently active.
One approach to KS scheduling is to use a control module that determines the order of KS activation on the basis of applicability and past use of the KSs.
As an extension of this idea, a separate blackboard system could select knowledge sources based upon explicit rule-based knowledge of the alternatives. This level of sophistication may, however, result in a slow response.
Abdul Rahim Ahmad
11
Blackboard System
Some degree of structure is imposed on the blackboard by dividing it into panels.
A knowledge source then only needs to look at a small number of panels rather than the whole blackboard.
Typically the blackboard panels correspond to different levels of analysis of the problem, progressing from detailed information to more abstract concepts.
Examples: Hearsay-II blackboard system for computerized understanding of
natural speech, the levels of analysis include those of syllable, word, and phrase.
Ultrasonic image interpretation using ARBS, the levels progress from raw signal data, via a description of the significant image features, to a description of the defects in the component
Abdul Rahim Ahmad
12
Blackboard System
Advantages of Blackboard System Many and varied sources of knowledge can participate in the
development of a solution to a problem.
Each knowledge source has access to the blackboard, so it can be applied as soon as it becomes appropriate. (opportunism, i.e., application of the right knowledge at the right time).
For problems involving large amounts of numerical processing, the characteristic style of incremental solution development is particularly efficient.
Different types of reasoning strategy (e.g., data- and goal-driven) can be mixed as appropriate in order to reach a solution.
Hypotheses can be posted onto the blackboard for testing by other knowledge sources. A complete test solution does not have to be built before deciding to modify or abandon the underlying hypothesis.
In the event that the system is unable to arrive at a complete solution to a problem, the partial solutions appearing on the
blackboard are available and may be of some value.
Abdul Rahim Ahmad
13
Genetic-fuzzy systems
Performance of a fuzzy system depends on the definitions of the fuzzy sets and on the fuzzy rules.
As these parameters can all be expressed numerically, it is possible to devise a system whereby they are learned automatically using genetic algorithms (GA).
A chromosome can be devised that represents the complete set of parameters for a given fuzzy system.
The cost function could then be defined as the total error when the fuzzy system is presented with a number of different inputs with known desired outputs.
Abdul Rahim Ahmad
14
Genetic-fuzzy systems
Fuzzy rules can be drawn up fairly easily, but defining the most suitable membership functions remains a difficult task.
A scheme by Karr is where all membership functions are triangular. The variables are constrained to lie within a fixed range, so the fuzzy
sets low and high are both right-angle triangles
The slope of these triangles can be altered by moving their intercepts on the abscissa, marked max1 and min4.
All intermediate fuzzy sets are assumed to have membership functions that are isosceles triangles. Each is defined by two points, maxi and mini, where i labels the fuzzy set.
The chromosome is then a list of all the points maxi and mini that determine the complete set of membership functions.
Abdul Rahim Ahmad
15
Genetic-fuzzy systems
Abdul Rahim Ahmad
16
9.3 Genetic-fuzzy systems
In this scheme, the genetic algorithm for parameter setting and the fuzzy system that uses those parameters are distinct and separate.
The parameters for a fuzzy system can also be learned using neural networks, where there is much closer integration between the neural network and the fuzzy system.
Abdul Rahim Ahmad
17
9.4 Neuro-fuzzy systems
A neuro-fuzzy system is a fuzzy system where the parameters of which are derived by a neural network learning technique.
Can also be viewed as a neural network that represents a fuzzy system.
The two views are equivalent : possible to express a neuro-fuzzy system in
either form.
Abdul Rahim Ahmad
18
Example
Consider the following fuzzy rules
These fuzzy rules and the corresponding membership functions can be represented by the neural network on the next page.
Abdul Rahim Ahmad
19
/* Rule 9.1f */IF temperature is high OR water level is highTHEN pressure is high/* Rule 9.2f */IF temperature is medium OR water level is mediumTHEN pressure is medium/* Rule 9.3f */IF temperature is low OR water level is lowTHEN pressure is low
Example (Contd.) Fuzzification: Give
membership value for low, medium, and high (using Single Layer Perceptron) for any given input values for temperature and water level .
Fuzzy rules: Input : the six membership values for temperature and water level. Outputs: 3 values for pressure.
Abdul Rahim Ahmad
20
Defuzzification: Combine these membership values to produce a defuzzified value for the output variable.
Example (Contd.)
The definitions of the fuzzy sets and the fuzzy rules are implicit in the connections and weights of the neuro-fuzzy network.
Using a suitable learning mechanism, the weights can be learned from a series of examples.
The network can then be used on previously unseen inputs to generate defuzzified output values.
In principle, the fuzzy sets and rules can be inferred from the network and run as a fuzzy rule-based system
Abdul Rahim Ahmad
21
9.5 Genetic-neural systems
A challenge in building a neural network is to choose the right architecture and the right learning parameters.
For a NN, a decision needs to be made on : the learning parameters
the necessary number of neurons
whether additional layers would be beneficial.
It is possible to use a genetic algorithm to optimize the network design. cost function might combine the RMS error with duration
of training.Abdul Rahim Ahmad
22
9.5 Genetic-neural systems
Supervised training of a NN involves
adjusting its weights until the output patterns obtained for a range of input patterns are close to the desired patterns.
The different network topologies use different training algorithms for achieving this weight adjustment, through back-propagation or errors.
However, it is also possible to use a genetic algorithm to train the network.
Abdul Rahim Ahmad
23
9.5 Genetic-neural systems
In the genetic-neural syatem: each gene represent a network weight, so that
a complete set of network weights is mapped onto an individual chromosome.
Each chromosome can be evaluated by testing a neural network with the corresponding weights against a series of test patterns.
A fitness value can be assigned according to the error, so that the weights represented by the fittest generated individual correspond to a trained neural network.Abdul
Rahim Ahmad
24
9.6 Clarifying and verifying NN In Neural networks, internals is a “black box” :
Ie: the inputs and outputs are significant to user but the weights is not.
In KBS, the internal state, (eg: value of a variable) does have meaning for the user.
Now, research are done to extract rule which are equivalent to the trained neural network from which they have been extracted including production rules and fuzzy rules. Example : in safety-critical systems.
Abdul Rahim Ahmad
25
Example: Safety-critical Systems
In safety-critical systems, cannot rely on output from a neural network without verification.
Need rules to verify consistency of a neural network output with its input.
Use of rules for verification implies that at least some of the domain knowledge can be expressed in rule form
One suggestion is to use an adjudicator module to decide whether a set of rules or a neural network is likely to provide the more reliable output for a given input.
Adjudicator need access to NN training data and could determine whether a NN would have to interpolate between, or extrapolate from, examples in the training.
Abdul Rahim Ahmad
26
Example: Safety-critical Systems
In safety-critical systems, cannot rely on output from a neural network without verification.
Need rules to verify consistency of a neural network output with its input.
Use of rules for verification implies that at least some of the domain knowledge can be expressed in rule form
One suggestion is to use an adjudicator module to decide whether a set of rules or a neural network is likely to provide the more reliable output for a given input.
Adjudicator need access to NN training data and could determine whether a NN would have to interpolate between, or extrapolate from, examples in the training.
Abdul Rahim Ahmad
27
Example: Safety-critical Systems
Neural networks are good at interpolation but poor at extrapolation. Thus, the adjudicator may call upon rules to handle the exceptional cases which would otherwise require a neural network to extrapolate from its training data.
If heuristic rules are also available for the less exceptional cases, then they could be used to provide an explanation for the neural network’s findings.
A supervisory rule-based module could dictate the training of a neural network deciding how many nodes are required adjusting the learning rate as training proceeds deciding when training should terminate.Abdul
Rahim Ahmad
28