|
|
|
 |
Note: Somehow, Newsletter #6 was lost. It was either not written or not sent properly. We apologize for any confusion
RiverSoftAVG
Newsletter #7 In this issue, the major
topic for discussion is the powerful new Expert System Wizard in the IECS EEP
(see "Getting Friendly with IECS v2.0 EEP" for more details).
There are also a couple of terrific tips: one about using the Combs method with
the EEP and the other about writing rules to create aggregate facts.
Finally, there are two new, great downloads you should know about. A reminder, this newsletter
and all previous ones are available for download from the web site at http://www.riversoftavg.com/articles_&_tips.htm
News:
10% Off on IECS v2.0/EEP ends March 31st, 2002 Tip: How
to implement the Combs method using IECS v2.0/EEP Fuzzy Logic to
prevent combinatorial explosion Tip: How
to write rules to collect information about some facts Free
Download: IECS/EEP Demo and Non-Source Version Due March 11-15. Free
Download: IECS/EEP v2.0 Expert System Wizard. Note: You received this
newsletter because you are an owner of a RiverSoftAVG product. If you
received this newsletter by mistake or if for any reason you wish to not receive
any future mailings from RiverSoftAVG, just reply to this message informing us
so. We apologize for any intrusion.
Before we start talking
about the wizard, we should note for our non-EEP owners that there is a version
of the wizard available for download from our web site. We converted the
wizard to an application which you may run stand-alone outside of Delphi;
unfortunately, unlike the wizard, it cannot fill in your TInferenceEngine
Component for you. It is intended only to show the features of the wizard.
You may download the application at How to start the Expert
System Wizard The Expert System Wizard is
a component editor. As such, it is accessed in Delphi by double-clicking
on the TInferenceEngine component. Alternatively, you can right click on
the TInferenceEngine component and Select the "Start Expert System
Wizard..." menu item. (If you are using the wizard application, just
start the application). How to use the Expert
System Wizard The Expert System Wizard
guides you in making an expert system by taking you, step-by-step, through the
various questions you need to answer to create one. This wizard works
similar to programs such as Tax Cut or Turbo Tax. The wizard asks you
questions and fills in the expert system based on your answers. In
addition to guiding you through the steps of creating an expert system, it also
provides some small expert systems for you. You may select these expert
systems to see how they are made and to "jump start" your own expert
system. The Expert System Wizard is a great tool for learning the
parts of an expert system (types, templates, fact, rules, etc) and for learning
the syntax of the IECS language. However, it cannot write the entire
expert system for you, you must provide the "intelligence" of the
expert system. When you start the expert
system, you are presented with a Welcome screen. Read this at your
leisure. Click the Next button a few times (or the Skip Button) to get
heart of the wizard. The Expert System Wizard
presents you with 6 steps for creating your expert system:
Selecting your Domain Your first step in creating
an expert system is to decide what you want to create it for. You must
decide the problem area, or domain, that you wish to reason about. There
are a few ready-made domains and expert systems available. You can also
load an expert system (IE File) from your hard disk and use the wizard to take
you through that expert system. If you started the wizard from the IDE,
your current expert system in the TInferenceEngine component is automatically
selected. For the purposes of our
walkthrough, we are going to use a pre-built expert system, Project Risk
Assessment, which is a fuzzy expert system for determining the risk of a project
based on various factors such as cost, manpower, etc. The Project Risk
Assessment expert system is under the Business Domain. Select the
"Business" domain and click Next. You will then see presented
some SubDomain choices. Select the "Project Risk Assessment
(Fuzzy)" and Click Next. The wizard will show you some text about
your selected domain. Click Next until you get back to the Agenda. Define Types The next four sections;
Define Types, Define Fact Templates, Define Rules, and Define Initial Facts; are
the real core of the wizard. In these four sections, you select and define
the constructs of your expert system. In this section, you start your
expert system by selecting and define the "simple" types, including
the fuzzy types, of your expert system. Defined Types are similar
to the enumerated types or integer sub-range types you would define in Object
Pascal (as opposed to Fact Templates, which are more similar to classes in
Object Pascal). DefTypes, while a unique construct of the IECS, borrow
concepts from (and provide some substitute features for) CLIPS deftemplate
slots. You can define the type (e.g., integer, string, fuzzy), numeric
range of the type (for integers, floats, and fuzzy), units, and allowed values
(similar to Object Pascal enumerations) or fuzzy terms. Types are not
necessary for creating expert systems. By default, all fact slots can be
ALL types (except fuzzy). However, deftypes provide enhanced type checking
and range checking. They are also required if you want to use fuzzy logic
in the IECS. If you click Next a few
times, you will get past the explanation pages and get to the "Select
Type" page. This page introduces you to a paradigm the wizard uses
again and again. On the left of the page, there is a pane detailing the
available types defined for this expert system. On the right is the
description and syntax to generate the selected type. If you were creating a new
expert system, the left pane would be blank except for the "Create New
Type..." item. To select types to include in your expert system, make
sure the items are checked. If you uncheck an item, it will NOT be in the
generated expert system. Note that unchecking a type can make other
constructs, such as fact templates or rules, invalid because they depend on the
unchecked type. You would need to uncheck those invalid constructs in the
Define Fact Templates, Define Rules section etc or recheck the type to generate
your expert system. Invalid constructs, whether because of syntax errors
or dependency errors, are displayed in the Left pane in RED. To edit a type, select the
type in the Left Pane and click Next. The following pages will take you
through all the decisions you need to make and the information you need to
provide to create that type. To create a new type, select the "Create
New Type..." item and click Next. Finally, to quit defining types,
click the Skip button to go to the next section of the wizard: Define Fact
Templates. Define Fact Templates The Define Fact Templates
section works exactly like the Define Types section. In this section, you
define the fact templates of your expert system. The fact template provides
the basic building blocks of your expert system. It defines a knowledge
unit, or fact type or fact class, in your domain. It serves as a
definition for a group of facts, similar to how a Pascal record definition or
class definition defines the structure for all variables of that type. The
deftemplate, as it is known in CLIPS, or fact template, defines the relation
name for a group of facts plus zero or more slots that encapsulate properties
about the class. To edit a fact template,
select the template in the Left Pane and click Next. The following pages
will take you through all the decisions you need to make and the information you
need to provide to create that fact template. To create a new fact
template, select the "Create New Fact Template..." item and click
Next. Finally, to quit defining templates, click the Skip button to go to
the next section of the wizard: Define Rules. Define Rules The Define Rules section is
where you put the intelligence in your expert system. It is where all the
hard work gets done. Rules define the
"code" of your expert system. They manipulate facts
("objects or data") to infer new facts or perform actions.
Unlike procedural languages, however, rules are always present and able to be
fired if there conditions are met. Rules are basically IF-THEN constructs
for reasoning about the domain. The inference engine matches rule
conditions (the IF portion) with facts in the expert system and flags rules
which have been matched (called a rule activation). After the inference
engine has determined all the rule matches, it selects one rule based on its
strategy and a rule's salience (priority) and fires that rule. The Define Rules section
works exactly like the previous two sections. You can edit existing rules
or create new rules. When you are finished defining rules, click the Skip
button to go to the Define Initial Facts section. Define Initial Facts This is the final section
in the wizard for defining expert system constructs. In this section, you
define the initial facts of your domain. Facts define pieces of
knowledge or data in your expert system. Facts define the world of your
expert system, i.e., the people in it, the things in it, etc. Rules
manipulate facts to infer new facts (or new knowledge) and to act. Facts
can be asserted into the world or retracted from the world. Facts are
based on fact templates. They consist of a relation name (the fact
template name) and zero or more slots and their associated values. For
example, the fact '(person (name Tom) (hair-color brown))' defines a person in
the world whose name is Tom and has brown hair color. Note that there are two
ways to define initial facts. The Expert System Wizard allows you to
define initial facts to be automatically asserted onto your fact list the first
time your expert system is run. However, subsequent executions will not
contain these facts. Fact Sets, or deffacts in CLIPS terminology, define a
group of facts that are asserted into your expert system every time the
inference engine is Reset. You can create fact sets by
dropping a TFactSet component on your form and clicking the component's Facts
property. After you are finished
editing or creating facts, click the Skip button to go to the wizard's final
section. Finish Up In this section, the wizard
allows you to review your choices. You get to see all the defined types,
fact templates, rules and initial facts in your system. The first page of
this section displays all the constructs in a tree. Constructs, which are
invalid, whether because of bad syntax or dependency problems, are highlighted
in red. To edit a construct (fact,
rule, etc), select the construct and click the Correct button. Only after
everything is correct are you taken to the final page. The final page
shows you what your expert system will look like. If you are running the
wizard application, the wizard ends here. You can select the expert system
text and copy it into a file for your use though. You may have to edit it
though, there may be some v2.0-isms in there which will make the expert system
invalid for v1.0 as written. If you are running the wizard from the IDE,
clicking Finish will fill your TInferenceEngine component with the new expert
system. Well, that was a short
overview of the IECS v2.0 Expert System Wizard. Obviously, such a massive
wizard can only be superficially explained here. We encourage you to
experiment with the wizard, either in the IDE or the download application.
If you have suggestions, please don't hesitate to write us. News: 10% Off
on IECS v2.0/EEP ends March 31st, 2002 IECS v2.0/EEP Full Version
(Source) $315
Until March 31st
$350 After March 31st IECS v2.0/EEP Upgrade
(Source) (Pre-August 1st 2001 Owners)
$157.50 Until March 31st $175 After
March 31st IECS v2.0/EEP Upgrade
(Source) (After-August 1st 2001 Owners)
$112.50 Until March 31st $125 After
March 31st IECS v2.0/EEP Crossgrade
(Non-Source to Source) (Pre-August 1st 2001 Owners)
$270 Until March 31st
$300 After March 31st IECS v2.0/EEP Crossgrade
(Non-Source to Source) (After-August 1st 2001 Owners)
$225 Until March 31st
$250 After March 31st IECS v2.0/EEP Full Version
(No Source) - Delphi 5/6
$180 Until March 31st
$200 After March 31st IECS v2.0/EEP Upgrade (No
Source) (Pre-August 1st 2001 Owners)
$90 Until March 31st
$100 After March 31st IECS v2.0/EEP Upgrade (No
Source) (After-August 1st 2001 Owners)
$76.50 Until March 31st $85 After
March 31st Note: IECS v2.0 is for
Delphi 5 and 6 only. Non-source code users will
also have a chance to buy v2.0/EEP at 10% off. The next EEP will be
available for non-source code users as well (see the Announcement below).
To reduce the development time (preparing non-source code versions is labor
intensive), that EEP will be the *only* EEP where non-source code is available
until the release of version 2.0. Tip: How to implement
the Combs method using IECS v2.0/EEP Fuzzy Logic to prevent combinatorial
explosion A handicap that expert
systems, especially fuzzy expert systems, face is what is called combinatorial
explosion. Combinatorial explosion occurs when the number of rules
increases exponentially with the increase in dependent variables (and their
values). A Boeing engineer, William E. Combs, has developed a method,
called oddly enough :-) the Combs method, for preventing combinatorial
explosion; this tip shows how to implement the Combs method with the IECS v2.0.
For the purposes of space, we will not explain in detail combinatorial explosion
and the Combs method. Please see "The Fuzzy Systems Handbook, 2nd
Edition" by Earl Cox or the "Imploding Combinatorial Explosion in a
Fuzzy System" gem by Michael Zarozinshi in the "Game Programming Gems
2" book, edited by Mark DeLoura for more details. Combinatorial Explosion of
Rules in Fuzzy Expert Systems Combinatorial explosion of
the number of rules occurs in Fuzzy Expert Systems as the number of dependent
variables, or fuzzy variables, increase the variables values, or fuzzy sets,
increase. For example, suppose you have written a fuzzy expert system
to determine the financial credit worthiness of an applicant. Your first
attempt at the fuzzy expert system uses two fuzzy variables: Income and Debt,
both of which may be either Low, Medium, or High. Your expert system also
has one solution fuzzy variable: Risk. Conventional thought for fuzzy
expert systems is that you must write a rule for every combination of your
dependent input variables: "if Income is Low and Debt is
Low then Risk is Medium," "if Income is High and Debt
is Low, then Risk is Low," etc. You must create a rule for every
combination of Income and Debt. A Fuzzy Expert System designed this way is
essentially what is called a Fuzzy Associative Memory, or FAM for your fuzzy
system. For two input variables (Income and Debt) and 3 fuzzy sets per
variable (Low, Medium, High), you would need 3 squared or 9 rules. If you
added just one extra fuzzy set value to Income and Debt, you would need 4
squared (16) rules. If you added another fuzzy variable, say Mental
Health, you would now have 4 Cubed (4^3=64) rules. As the number of fuzzy
variables or fuzzy sets increase, the number of rules increases exponentially: Number
of
Number of
Number of Fuzzy
Variables Fuzzy Sets/Var Rules
2 3
9
2 4
16 3 3
27 3 4
64 4 4
256 5 4
1024 10 4
1048576 In practice, using hedges
and smart rule writing, you don't need a separate rule for each combination.
However, the concern is valid, rules do increase exponentially. What can
we do? Well, William Combs
developed a method to reduce this exponential growth to linear growth.
Instead of 10 Fuzzy Variables with 4 Fuzzy Sets Each having over a million
rules, the Combs method makes you end up with 40 rules! Basically, the
Combs method realizes that the logical proposition (p and q) then r is
equivalent to (p then r) OR (q then r), where p, q, and r represent terms such
as "Income is Low," "Debt is Low," and "Risk is
Low" respectively. What does this mean in
practice? Instead of writing rules for every combination of input
variables and fuzzy set values, you just write rules for each input variable by
itself: "if Income is Low,
then Risk is High" "if Income is Medium,
then Risk is Medium" "if Income is High,
then Risk is Low" "if Income is Obscene,
then Risk is Low" With the IECS, you just
need to rewrite the rules... everything else happens automatically. Not only does the Combs
method reduce the number of rules, it also reduces maintenance. Suppose
you did want to add a new input variable to your expert system. With the
traditional technique, you would have to modify ALL of your existing rules as
well as adding new ones! With the Combs method, you just add rules for the
new input variable. What's the catch? Ok,
it does seem to be too good to be true. But in this case, there is little
penalty for using the Combs method. With a fuzzy expert system, the Combs
method does give slightly different results than the traditional
method. This is more a matter of how fuzzy inference engines work than the
method itself. When you AND variables together, the Fuzzy Inference Engine
uses the MINIMUM pattern match strength of all the variables. A rule such
as "if Income is Low AND Debt is High..." may match with facts so that
Income matches with a 17% strength and Debt matches with a 80% strength.
The Fuzzy Inference Engine activates this rule with the minimum of the two
strengths: 17%. This activation strength modifies any assertions made in
the rule, e.g., assert Risk is High with a cutoff of 17%. The Combs method
avoids this minimization, it just ORs the results together. The Combs method provides
great advantages. It reduces the number of rules and maintenance of those
rules. Plus, it provides faster expert systems! The Inference Engine
does not have to perform so much work matching facts with potentially thousands
of rules. You owe it to yourself to consider using the Combs method with
your expert systems. Tip: How to write
rules to collect information about some facts Occasionally in your expert
systems, you may find yourself needing to look at all facts of one type and find
some aggregate information. For example, you may have a list of customer
facts and you want to find the average income of your customers.
Unfortunately, expert systems do not give an easy way for finding this aggregate
information. This tip will illustrate one way to write some rules to find
the average of a fact slot, the average income in all the customer facts on
your fact list. Let's fire up the Advanced Console Editor to create our
expert system. First, we need to define
the fact templates we are working with. Of course, there is the customer
fact template. For our purposes, we are going to ignore all slots and just
ensure we have an income slot. To calculate an average, we need to sum all
the customer ages and then divide by the number of customers. Therefore,
we need an intermediate fact to hold the sum and fact count. Here are the
two fact templates we came up with (deftemplate
customer Now, you may just be
tempted to write a rule that matches with a customer fact and modifies the
customer-all fact: (defrule
build-sum This rule will have one
activation for every customer fact in the fact list. However, there is a
problem. As written, this rule causes an infinite loop! Every time
the customer-all fact is modified, this rule gets reactivated for the new
customer-all fact. What can we do? Well, we need to prevent
the reactivation. The CLIPS way is to create another intermediate
fact, say add-sum. Every customer fact would assert one add-sum fact.
This add-sum fact would not cause reactivation of the summing rule. Then,
another rule would add the add-sum facts to make customer-all. Here are
the modified rules: (defrule
build-sum (defrule
build-sum-end Finally, we need to create
a rule to calculate the actual average. For our purpose, we will just
printout the average. The only tricky part about this rule is that we
don't want it to activate on the customer-all fact until every other rule has
finished firing. To do this, we just declare the salience (or priority) to
a negative number; remember that salience is different than priority in that lower
values have LESS priority. Since all the other rules have the default
salience of 0, this averaging rule will fire last: (defrule
calc-average (declare
(salience -10))
(customer-all (count ?count) (sum ?total) ) Voila! We're done.
We just need to add the customer facts and the customer-all facts to test our
expert system. The fact set (deffacts) for that is below: (deffacts
myfacts Reset the expert system and
run! Next time, we will show you how to create a user function to do the
same function.
The Inference Engine
Component Suite version 2.0 Early Experience Program will release a new version
next week, hopefully on the Ides of March :-). This release will have bug
fixes as well as help updates. The big news is that next week a demo
version AND a non-source code version will also be released. To download
the demo, check back at http://www.riversoftavg.com/downloads.htm at
the end of next week. Non-source code users will be able to upgrade to the
new version at the same time. And now through March 31st, the upgrade will
be available at 10% off.
The Inference Engine
Component Suite version 2.0 Expert System Wizard is available as a stand-alone
application from our downloads page. This application allows you to try
out the wizard for yourself if you don't have the EEP. Though it cannot
generate the expert system automatically for you, you can copy the expert system
to the clipboard to try and use with the Advanced Console Demo. Our
article above, Getting Friendly with IECS v2.0 EEP, also talks about the Expert
System Wizard. |
Send mail to
webmasterNO@SPAMRiverSoftAVG.com with questions or comments about this web
site.
|
