Title: ICT619 Intelligent Systems
1- ICT619 Administrative Notices
- Note No class Thursday 20th Sept.
- If you haven't already, please submit your 1
page project proposal as soon as possible
2ICT619 Intelligent SystemsTopic 3 Fuzzy Systems
3Fuzzy Systems
- PART A
- Introduction
- Applications
- Fuzzy sets and fuzzy logic
- Probability and fuzzy logic
- Fuzzy reasoning
- Design of a fuzzy controller
- PART B
- Building fuzzy systems
- Advantages and limitations of fuzzy systems
- Case Studies
4Building fuzzy systemsPrincipal issues
- Validation of proposed application
- Fuzzy set and data representation
- Fuzzy set hedges
- Dealing with Boolean variables
- Uncertain and noisy data
- Fuzzy system design methodology
5Validation of proposed application
- Involves comparing the problem attributes with
those of a typical fuzzy model - Characteristics of a fuzzy model
- Intrinsically imprecise parameters (control
variables) - The domain expert expresses heuristics using
vague linguistic expressions for example,
relatively high profit margin, small-to-medium
sized companies. - Draws on knowledge from multiple experts
-
6Validation of proposed application (contd)
- Characteristics of a fuzzy model (contd)
- Experts express apparently conflicting opinions.
For example, - Our price must be high Company financial
expert - Our price must be low Company marketing and
sales expert. - Complex, poorly understood, nonlinear problems
- No mathematical model exists to represent it.
- Too difficult for traditional expert systems,
mathematical and statistical approaches.
7Fuzzy set and data representation
- Problem variables define data used and produced
by the system - Variables decomposed into one or more fuzzy sets,
with each set describing some range of the
variables values - Each domain value (for example, profit figures in
dollars) must belong to at least one fuzzy set
8Fuzzy set and data representation(contd)
There should be between 25 and 50 overlaps
between adjacent fuzzy sets.
9Designing and eliciting fuzzy sets
- Knowledge acquisition
- Sources include domain experts, articles,
procedure manuals, or other documents - All relevant objects (variables) and their
properties (fuzzy sets) are extracted - Derivation of fuzzy sets
- Voting by multiple experts
- Statistical analysis of domain data
- Data mining techniques
10Derivation of fuzzy sets
- Voting by multiple experts
- A fuzzy sets membership function can be based on
cumulative frequency of votes by experts - Eg, Membership value of 170 cm in tall determined
by how many regard 170 cm as tall - Experts must agree on the underlying domain
tall for a man? tall for a basket ball player?
11Derivation of fuzzy sets (contd)
- Statistical analysis of domain data
- For example, data are normally distributed, mean
and standard deviation are known - But data in financial analysis, marketing, risk
assessment, project management and so on are
seldom normally distributed and can be subject to
sudden changes - Data mining techniques
- Used to decompose underlying variables into
arbitrary collections of fuzzy sets
12Fuzzy set hedges
- Hedges are modifiers of fuzzy set membership
functions. - Act like adjectives and adverbs in English, such
as "very" or "often" - Allow a closer and more intuitive modelling of
the underlying knowledge as expressed in language - Defining hedges can be a subjective process and
may vary from one problem domain to another
13Fuzzy set hedges (contd)
- Some examples
- Very, almost, about, roughly, quite, more or
less, somewhat, rather - Membership very of variable x (membership of
variable x)2 - If income is low to a degree of 0.8, then income
will be very low to a degree of (0.8)2 0.64. - Membership more or less of variable x
?(membership of variable x) - Hedges reduce the number of fuzzy sets that need
to be created - Hedges also increase the readability of rules and
consequently ease maintenance.
14Dealing with Boolean variables
- Problem with Boolean expressions
- Boolean variables in the antecedents of rules may
saturate consequent fuzzy space, and pre-empt all
the other rules - Example
- Rules from a vehicle underwriting risk
assessment model - IF age is young THEN risk is increased
- IF distance_to_work is far THEN risk is slightly
increased - IF accidents_in_last2years is excessive THEN risk
is very increased - IF drink_driving_convictions is significant THEN
risk is totally increased - Now consider the addition of the following rule
- IF sexmale THEN risk is increased
- may cause the consequent truth value to evaluate
to 1.0 if sex happens to be male
15Dealing with Boolean variables(contd)
- Boolean filters
- A Boolean filter can be added to the antecedent
of a rule as a qualifying proposition to avoid
saturation - IF sexmale AND age is young THEN risk is
highly increased - If the expression sexmale is true (value
1), and the truth of age is young is say 0.38,
the consequent will be assigned the truth value
minimum(1.0, 0.38) or 0.38 (using the min-max
rule). - When the expression sexmale is false (value
0), - the minimum is zero and the rule is not executed
(which is the original intent). - Explicit truth values for Boolean propositions
- For example the rule
- IF sexmale.60 THEN risk is increased
- agrees with the intuition overall risk is
increased independent of any other factors if the
driver is male
16Uncertain and noisy data
- Data may be noisy or uncertain due to
- imprecise measurement processes or
- lack of confidence in the data source
- May be handled by associating a fuzzy set with
the data also - As an example, one of the input variables in the
rule - IF income is moderately_high THEN credit_risk
is medium - is the income stated by the candidate
17Uncertain and noisy data (contd)
- The assessor may associate a degree of
confidence, say 0.8, with the income - the lower the confidence value, the broader
the fuzzy numbers bell curve - The truth value of the antecedent is found by
finding the point of maximum correlation between
the antecedent fuzzy set moderately_high and the
fuzzy set associated with the stated income
(135,000 in this example)
Confidence set
18Uncertain and noisy data (contd)
- Application of the rule
- IF income is moderately_high THEN credit_risk
is medium - with cf0.8 for variable income will thus
truncate the fuzzy space for medium credit risk
to the level 0.7.
19Fuzzy system design methodology
- The basic steps
- Validate application
- Define functional and operational characteristics
- Decompose variables into fuzzy sets
- Write rules to describe behaviour
- Define defuzzification methods
- Define performance metrics
- Test system
- Evaluate and tune system
20Definition of fuzzy sets
- No formal procedure for designing these
functions - Practical experience and intuition used
- Experimental evidence indicates a high tolerance
to fuzzy shapes - In control applications, the fuzzy sets are
usually trapezoidal or triangular - Bell-shaped fuzzy sets may also be used instead
of triangles - Sigmoids (S-curves) and linear surfaces are also
used in information systems and applications in
social sciences.
21Writing rules
- Rules are usually written in the form
- IF x is S THEN y is T
- where x and y are linguistic variables and S and
T are fuzzy sets - In this rule, x is a control (input) variable and
y is the solution (output) variable - Any unconditional rules of the form
- x is S
- is also entered.
- Such rules act to constrain the shape of the
solution fuzzy set.
22Review of rules and addition of any hedges
- If the fuzzy system supports hedges, the rule set
should be examined to write any hedged rules - These rules generally capture extreme behaviour
patterns, eg, - IF x is very S THEN y is very T
- Addition of any alpha cuts to individual rules
- An alpha cut establishes the minimum truth
threshold for a rule - For example, given the rule
- (Alpha0.2) IF age is young THEN risk is high
- if the truth of the proposition age is young
falls below 0.2, this rule will not fire - Prevents production of residual fuzzy spaces
with very small but finite truth values
23Entering rule execution weights
- Importance of rules can be indicated by giving a
weight multiplier. For example, - (Weight0.8) IF age is young THEN risk is high
- Definition of defuzzification method
- Defuzzification selects the expected value of a
solution variable from the aggregated fuzzy
region. - The most popular method is the centre of gravity
or centroid as it gives a smoother transition
from one observation to the next.
24Advantages of fuzzy systems
- Important attributes used by organisations in
evaluating systems - Mean-time-between-failure (MTBF)
- Mean-time-to-repair (MTTR)
- Extensibility of existing systems
- Agreement with real-world processes
- Understandability
- Business intelligent systems based on fuzzy logic
show advantages in all these aspects
25Advantages of fuzzy systems (contd)
- Some benefits of fuzzy systems are
- Ability to model highly complex business problems
- Conventional expert systems failed to grow in
business applications mainly due to the complex
nature of many business problems - Fuzzy systems are highly suited for modelling
computationally complex non-linear problems which
are poorly understood - Efficient knowledge base
- Fuzzy rule based systems require fewer rules and
execute faster than conventional rule based
systems. - A single rule applies by different degrees to a
variety of situations.
26Advantages of fuzzy systems (contd)
- Improved cognitive modelling of expert systems
- Specialized knowledge is often best expressed
using imprecise terms. Conventional expert
systems force experts to express rules using
crisp boundaries - This contributes to poor performance in many
systems. - Ability to model systems involving multiple
experts - Since rules are applied in parallel rather than
in sequence, - two rules reflecting two different opinions can
both contribute to the final outcome.
27Advantages of fuzzy systems (contd)
- Reduced model complexity
- Fuzzy systems use fewer and the rules are closer
to the way knowledge is expressed in natural
language. These factors greatly improve overall
MTTR. - The understandability and ease of maintenance
also leads to improved MTBF. - Improved handling of uncertainty
- Fuzzy systems provide a better, more consistent
and mathematically sound method of handling
uncertainties - Example
- IF height is tall THEN weight is heavy
- A given height will produce an estimated weight
with its degree of membership in the fuzzy set
heavy. - An expert system would contain many rules, like
- IF height is gt 135 and height is lt 165 THEN
weight is 75, CF0.8
28Limitations of fuzzy intelligent systems
- Lack of learning capability
- Hybrid, particularly neurofuzzy systems, attempt
to remedy this - Determining and tuning membership functions
- Is not an easy task
- Difficult to say how many are really required
- Requires extensive testing and fine tuning
- Though easier to design and prototype than
conventional systems, fuzzy systems require more
fine tuning before they are operational. - Bias among traditionalists, especially in Western
countries, for crisp logic based systems - Mistrust of fuzzy logic
29Case Studies
- Case 1 Quality control and monitoring (Dhar
Stein 1997 p.203-210) - German superstore chain Kaufhof
- Manufacturers supply directly to 16 warehouses.
- Problem
- 112,000 deliveries from manufacturers each day
- About 1 have a problem (quality, type, number,
breakages, etc.) - Checking all items individually is very labour
intensive
30Case Study 1 (contd)
- Solution goal
- A continual monitoring system for identifying
shipments with high risk - Constraints
- Should concentrate on riskier shipments
- Be flexible enough to deal with new products,
suppliers and quality control policies - Be fast enough not to disrupt the workflow at the
warehouse
31Case Study 1 (contd)
- Relevant issues
- Factors affecting shipments
- The nature of items being delivered
- Past performance of the supplier
- Recent performance trends of the supplier
- The size of the order
- Factors interact in complex ways
- Evaluation
- Error rate of the brute-force system to be used
as a benchmark for evaluating new system.
32Case Study 1 (contd)
- Possible solutions
- ANN and statistical analysis solutions ruled out
Reason No data available relating input
(shipment attributes) to output (shipment errors) - Expert systems considered Reason Experts able
to articulate criteria for wanting to inspect
shipments
33Case Study 1 (contd)
- The fuzzy system solution - Invent/W
- Fuzzy system solution chosen due to ability to
- model processes with varying degrees of truth and
- deal with the interaction of continuously varying
variables - Information in shipment barcode (eg, product,
supplier) used by Invent/W to gather input
variables based on - historical data
- trend data
- product type data etc
34Case Study 1 (contd)
- Invent/W uses
- Inputs from Kaufhof stores in the inventory
system - ancillary databases
- special tables used for building risk profiles
- Analyses these data and produces the solution
variable - a score between 0 and 100 - The lower the number, the more the shipment
should be inspected
35Case Study 1 (contd)
- The Invent/W rule base
- 164 rules in the system
- Rules consider suppliers' recent performance
- A typical rule
- IF recent_shipment_problem is high AND
item_shipment_risk is high THEN risk is high - System integrated into a COBOL-based inventory
management system, which runs on Kaufhofs
mainframe computer - A Windows based graphical environment enable
users to easily define rules and the shapes of
fuzzy sets
36Case Study 1 (contd)
- Invent/W performance test
- All shipments inspected regardless of score
output by Invent/W - In first four months of testing, 98.5 of all
erroneous shipments identified correctly - Currently performing at about 99.5 accuracy
- Shipment inspection capacity required reduced by
more than 50.
37Case Study 2 Bond rating (McNeill Thro 1994)
- A neuro-fuzzy system for rating the investment
safety of bonds - Developed by Fujitsu in Japan
- The system uses 10 fuzzy rules
- Fuzzy membership functions and the rule weights
adjusted using a neural network. - Rules cover the inputs and the financial
condition of the company issuing the bonds
See http//www.fuzzysys.com/books/FLLib/FUZZYPDF
/FUZZYLOG.PDF
38Case Study 2 (contd)
- Financial condition of company issuing bonds
determined by - Ordinary profit (with membership functions for
fuzzy sets large, medium, small) - Owned capital (large, small)
- Interest coverage ratio (high, low)
- Long-term loan ratio (low, high)
- Owned capital ratio (low)
- Outputs (bond ratings) are high, medium, and low
39Case Study 2 (contd)
- The system uses two classes of rules
- Basic rules
- receive an initial weighting of 1, are related to
ordinary profits - IF ordinary profit is large THEN rating is high
- IF ordinary profit is medium, THEN rating is
medium - IF ordinary profit is small, THEN rating is low
- Auxiliary rules
40Case Study 2 (contd)
- Auxiliary rules
- Cover the other inputs and receive an initial
weighting of 0.2 - IF owned capital is large, THEN rating is high
- IF owned capital is small, THEN rating is small
- IF interest coverage ratio is high, THEN rating
is high - IF interest coverage ratio is low, THEN rating is
low - IF long-term loan ratio is high, THEN rating is
high - IF long-term loan ratio is low, THEN rating is
low - IF owned capital ratio is low, THEN rating is
low - Initial weight values associated with rules
either increased or decreased after learning by
the ANN.
41REFERENCES
- Cox, E., The Fuzzy Systems Handbook, AP
Professional, San Diego 1999. - Dhar, V., Stein, R., Seven Methods for
Transforming Corporate Data into Business
Intelligence., Prentice Hall 1997, pp. 126-148,
203-210. - Mcneill, F., Thro, E., Fuzzy Logic a Practical
Approach, AP Professional, Boston 1994. - Munakata, T., Jani, Y., Fuzzy Systems an
Overview, Communications of the ACM, Vol.37,
No.3, 1994, pp.69-76. - Medsker,L., Hybrid Intelligent Systems, Kluwer
Academic Press, Boston 1995. - Negnevitsky, M. Artificial Intelligence A Guide
to Intelligent Systems, Addison-Wesley 2005.
Sangalli, A., The Importance of being Fuzzy,
Princeton University Press, 1998. - Zahedi, F., Intelligent Systems for Business,
Wadsworth Publishing, Belmont, California, 1993.
42ICT619 Student Projects from 2004
Iris dataset - ANN (MATLAB)
SW project cost estimation - ANN
Social skills assessment for kids - ES (EXpert2go)
HDD problem diagnosis - ES
PC troubleshooting - ES
Display troubleshooting - ES (CLIPS)
NW troubleshooting - ES (Exper2go)
Char recognition - ANN (JOONE)
Wine recommender - ES(Exp2Go)