Knowledge-Based Systems Expert Systemscobweb.cs.uga.edu/~potter/kbs/VaTechECE4524.pdf · MYCIN uses...
-
Upload
trinhkhanh -
Category
Documents
-
view
221 -
download
0
Transcript of Knowledge-Based Systems Expert Systemscobweb.cs.uga.edu/~potter/kbs/VaTechECE4524.pdf · MYCIN uses...
9expert.doc EE 4524 Page 1
ECpE 4524
Knowledge-BasedSystems
often called
Expert Systems
9expert.doc EE 4524 Page 2
Knowledge-based systems(textbook, chapter 20)
Goal:Try to solve the kinds of problemsthat normally require human experts
Typical examples:medical diagnosis, financial analysis,factory production scheduling
Why study knowledge-based systems?To understand human reasoning
methodsHuman experts tend to take
vacations, get hired by othercompanies, ask for raises, retire,become ill, die, . . .
Lots of commercial successes!
9expert.doc EE 4524 Page 3
Expert system overview:
KNOWLEDGEBASE
REASONINGMECHANISM
Problemdescription
Analysis andjustification
The knowledge base . . .● contains "domain knowledge,"
normally provided by human experts● is typically very specialized for a
particular problem domain● is often encoded as IF-THEN rules● may incorporate heuristics or
probabilities● is a valuable commodity
9expert.doc EE 4524 Page 4
Building, validating, and maintaininga knowledge base is a skill (art)called knowledge engineering
The reasoning mechanism . . .● takes descriptions from the user
about the problem to be solved● requests additional information
from the user as needed● interprets the knowledge base to
make inferences, drawconclusions, and ultimately giveadvice
● explains its reasoning to the user(how were the conclusionsreached?)
● is sometimes called aninference engine
9expert.doc EE 4524 Page 5
An example:
PUFF (1979)Pulmonary function analysis
Physician refers patient topulmonary testing lab
Patient inhales/exhales throughtube connected to computerizedinstrument which measures flowrates and air volumes
PUFF accepts this data along withauxiliary data (age, sex,smoking history), and printsdiagnosis in English
Now used on a routine basis (?)
9expert.doc EE 4524 Page 6
Building PUFF's knowledge base:A knowledge engineer sat downwith an expert pulmonaryphysiologist at the PacificMedical Center in SanFrancisco and developed rules(64 in all)
IF 1. ----------2. ----------3. ----------. . .
THEN 1. ----------2. ----------. . .
(A more recent version had about400 rules.)
9expert.doc EE 4524 Page 7
Example PUFF rule:
RULE31IF:
1. The severity of obstructive airwaysdisease of the patient is greater than or equalto mild, and2. the degree of diffusion defect of thepatient is greater than or equal to mild, and3. the TLC observed/predicted of the patientis greater than or equal to 110, and4. the observed/predicted difference inRV/TLC of the patient is greater than orequal to 10
THEN:1. There is strongly suggestive evidence(0.9) that the subtype of obstructive airwaysdisease is emphysema, and2. It is definite (1.0) that "OAD, DiffusionDefect, elevated TLC, and elevated RVtogether indicate emphysema" is one of thefindings
9expert.doc EE 4524 Page 8
Sample PUFF session:
PATIENT DATA:The degree of dyspnea: MODERATELY-SEVEREThe severity of coughing: MILDSputum production: MODERATELY-SEVEREThe number of pack-years of smoking: 48Referal diagnosis: BRONCHITISIVC/IVC-predicted: 80RV/RV-predicted: 191FVC/FVC-predicted: 87TLC observed/predicted: 127Predicted FEV1/FVC: 83TLC(DLCO) observed/predicted: 83FEV1/FVC ratio: 50RV/TLC observed/predicted: 21MMF/MMF-predicted: 19DLCO/DLCO-predicted: 48The slope of (F50obs-F25obs)/FVCobs: 19
DEGREE OF OBSTRUCTIVE AIRWAYS DISEASE:OAD degree by slope: MODERATELY-SEVERE 700OAD degree by MMF: SEVERE 900OAD degree by FEV1: MODERATELY-SEVERE 700Final OAD degree: MODERATELY-SEVERE 910
SEVERE 900
9expert.doc EE 4524 Page 9
INTERPRETATION:Obstruction is indicated by curvature of the flow-volume
loop.Forced Vital Capacity is normal and peak flow rates are
reduced, indicating severe airway obstruction.Change in expired flow rates following bronchodilation
shows that there is reversibility of airwayobstruction.
Elevated lung volumes indicate overinflation.Air trapping is indicated by the elevated difference
between observed and predicted RV/TLC ratios.Airway obstruction is consistent with the patient's
smoking history.The airway obstruction accounts for the patient's dyspnea.Although bronchodilators were not useful in this one
case, prolonged use may prove to be beneficial.Obstructive Airways Disease of mixed types.
9expert.doc EE 4524 Page 10
How were the rules produced?100 cases (previously diagnosed
patients) were selected
The cases were chosen to span thevariety of known disease states
The pulmonary function expert posedhypothetical rules for diagnosingthe illness
The knowledge engineer encodedthe rules (in LISP) and testedthem with the test cases.
The expert reviewed the test resultsand modified or added rules tohandle the cases that wereincorrectly diagnosed
Looping continued until the expertwas satisfied
9expert.doc EE 4524 Page 11
How to test PUFF's performance?150 additional different cases were
analyzed1) by human experts and2) by PUFF
The diagnoses were compared:90% matched to same degree ofseverity100% matched to within onedegree of severity
Effort:50 hours by the expert400 hours by the knowledgeengineer
The 64 rules were "popped into" anexisting expert system
9expert.doc EE 4524 Page 12
OBSERVATIONSHuman experts are often unaware of
how they reach conclusionsThe expert usually knows more than
he/she is aware of knowingThe knowledge brought to bear by the
expert is often experiential, heuristic,and uncertain
General problem-solvers (domain-independent) are too weak forbuilding real-world, high-performance systems
The behavior of the best problem-solvers (humans) is weak andshallow except in areas ofspecialization
Expertise in one specialization areausually does not transfer well to otherareas
9expert.doc EE 4524 Page 13
Recall weak vs. strong methods:
Weak methodsdomain-independent, general-
purpose(example: GPS)
Strong methodsdomain-specific, knowledge-rich(examples: knowledge-based
systems)
9expert.doc EE 4524 Page 14
Example expert systemsMedicineMYCIN (1976)Identification of bacteria in blood and
urine samples; prescription ofantibiotics
INTERNIST / CADUCEUS (1970s /1984)
Diagnosis of majority of diseases in fieldof internal medicine
PUFF (1979)Interpretation of respiratory tests for
diagnosis of pulmonary disordersBABY (19??)Patient monitoring in a newborn
intensive care unitQMR (1988) (Quick Medical Record)Assists physicians in diagnosis of over
4000 disease manifestations (usesthe INTERNIST knowledge base)
9expert.doc EE 4524 Page 15
CHEMISTRYDENDRAL (1960s and 1970s)Identification of molecular structure
of organic compoundsCRYSALIS (19??)Interpretation of electron density
maps in protein crystallographyMOLGEN (19??)Planning DNA-manipulation
experiments in moleculargenetics
AGRICULTUREPLANT/dsDiagnosing diseases in soybeansPLANT/cdDiagnosing cutworm damage in
corn
9expert.doc EE 4524 Page 16
OTHERSPROSPECTOR (1978)Provides advice on mineral prospectingMACSYMA (1968 - present)Symbolic solutions to mathematical
problemsR1 / XCON (1982)Configures VAX computer systemsGATES (1988)Used by TWA at JFK airport to assist
ground controllers in assigning gatesto arriving and departing flights
DESIGN ADVISOR (1989)Critiques IC designsTOP SECRET (1989)Decide the correct security classification
to give a nuclear weapons document
9expert.doc EE 4524 Page 17
DENDRALFeigenbaum (1960s and 70s)
One of the first expert systems
Identifies of molecular structure oforganic compounds
Uses mass spectrogram andnuclear magnetic resonance(NMR) data
9expert.doc EE 4524 Page 18
MYCIN (a precursor to PUFF)(textbook, Section 8.2)
Shortliffe, 1976 (Stanford, in Interlisp)
MYCIN is possibly the best knownexpert system that has beendeveloped
MYCIN can diagnose bacterialinfections and recommendtreatment
MYCIN was designed solely as aresearch effort (never used inpractice?);it served as a model for muchsubsequent expert-system work
9expert.doc EE 4524 Page 19
MYCIN uses IF-THEN rules
Each rule has a certainty factorassociated with it(a measure of the confidenceassociated with the rule's conclusion)
exampleRULE50IF 1) the infection is primary-bacteremia,
and2) the site of the culture is one of thesterile sites, and3) the suspected portal of entry of theorganism is the gastrointestinal tract
THENthere is suggestive evidence (0.7) thatthe identity of the organism isbacteroides
9expert.doc EE 4524 Page 20
in LISP:PREMISE:(AND
(SAME CNTXT INFECTPRIMARY-BACTEREMIA)(MEMBF CNTXT SITESTERILISITES)(SAME CNTXT PORTAL GI) )
ACTION:(CONCLUDE CNTXT IDENT
BACTEROIDES TALLY 0.7)
Each rule:premise ⇒ action
Each premise:logical conjunction of clauses
Each clause:<predicate> <object> <attribute> <value>
9expert.doc EE 4524 Page 21
MYCIN used a front-end known as
TEIRESIAS(textbook, section 20.3)
Davis, 1977, 1982
TEIRESIAS is a knowledge-acquisition manager; it providesexplanations of how conclusionsare reached
9expert.doc EE 4524 Page 22
Example TEIRESIAS rule(actually a meta-rule)
IF1) the patient is a compromisedhost, and2) there are rules that mentionin their premise pseudomonas,and3) there are rules that mentionin their premise klebsiellas
THENthere is suggestive evidence(0.4) that the former should bedone before the latter
9expert.doc EE 4524 Page 23
Example MYCIN / TEIRESIASsession:
1) Patient's name:** Fred Smith2) Sex** male3) Age:** 554) Have you been able to obtain positive cultures
from a site at which Fred Smith has aninfection?
** yes5) What is the infection?** Primary-bacteremia6) Please give the date and approximate time
when signs or symptoms of the primary-bacteremia (infection-1) appeared.
** May 5, 19807) From what site was the specimen for culture-1
taken?** Blood8) Please give the date and time when this blood
culture (culture-1) was obtained.** May 9, 1980
9expert.doc EE 4524 Page 24
9) Enter the identity of organism-1.** unknown10) Is organism-1 a rod, or coccus (etc.)?** rod11) The gram stain of organism-1:** grammeg12) Have there been positive cultures from sites
which are NOT associated with infections aboutwhich you are seeking advice?
** no13) Did organism-1 grow aerobically?** yes (0.8)14) Did organism-1 grow anerobically?** yes15) What is the suspected portal of entry of
organism-1 into the sterile site?** GI (0.6). . . etc.
9expert.doc EE 4524 Page 25
My therapy recommendation will be based on thefollowing probable infection(s) and potentialcausative organism(s):
INFECTION-1 is PRIMARY-BACTEREMIA<Item 1> Organism-1 may be PSEUDOMONAS-
AERUGINOS<Item 2> Organism-1 may be KLEBSIELLA-
PNEUMONIAE<Item 3> Organism-1 may be E. COLI<Item 4> Organism-1 may be BACTEROIDES-
FRAGILIS<Item 5> Organism-1 may be ENTEROBACTER<Item 6> Organism-1 may be PROTEUS-NON-
MIRABILIS
46) On a scale of 0 to 4, where higher numbersindicate increasing severity, how would you rateFred Smith's degree of sickness?
** 347) Does Fred Smith have a clinically significant
allergic reaction to any antimicrobial agent?** no48) Fred Smith's weight in kilograms:** 70
9expert.doc EE 4524 Page 26
[Rec 1] My preferred therapy recommendation is asfollows:
In order to cover for Items <1 2 3 5 6>Give: GENTAMICINDose: 119 mg q8h IV [or IM] for 10 daysComments: Modify dose in renal failure
In order to cover for Item <4>Give: CLINDAMYCINDose: 595 mg q6h IV [or IM] for 14 daysComments: If diarrhea or other GI symptomsdevelop, patient should be evaluated forpossible pseudomembranous colitis
9expert.doc EE 4524 Page 27
Do you wish to see the next-choice therapy?** yesWhich, if any, of the above drugs do you wish
eliminated from consideration?** Clindamycin
[Rec 2]In order to cover for Items <2 3 4 5 6>
Give: CHLORAMPHENICOLDose: 1 mg q6h IV for 14 daysComments: Monitor patient's white count
In order to cover for Item <1>Give: GENTAMICINDose: 119 mg q8h IV [or IM] for 10 daysComments: Modify dose in renal failure.
9expert.doc EE 4524 Page 28
The ability to explain the reasoningpath is important
(MYCIN:) Is the patient's illness withorganism-1 a hospital-acquired infection?
(USER:) ** why(TEIRESIAS:) (WHY is it important to determine
whether or not the infection with organism-1 wasacquired while the patient was hospitalized?)
[1.0] This will aid in determining the category oforganism-1. It has already been establishedthat
[1.1] the gram stain of organism-1 is gramneg,and
[1.2] the morphology of organism-1 is rod, and[1.3] the aerobicity of organism-1 is faculativeTherefore, if[1.4] the infection with organism-1 was not acquired
while the patient was hospitalizedthen there is strongly suggestive evidence (0.8) that
the category of organism-1 isenterobacteriaceae.
9expert.doc EE 4524 Page 29
Also: there is weakly suggestive evidence (0.1) thatthe identify of organism-1 is pseudomonos-aeruginosa [RULE37].
(USER:) how 1.3(TEIRESIAS:) (HOW was it established that the
aerobicity of organism-1 is faculative?)The following were used:[3.1] RULE27 indicated there is strongly
suggestive evidence (0.8) that the aerobicity oforganism-1 is faculative.
[3.2] RULE27 indicated there is weaklysuggestive evidence (0.2) that the aerobicity oforganism-1 is anerobic
Since this gave a cumulative CF of (0.8) forfaculative, and (0.2) for anerobic, it has beenestablished that the aerobicity of organism-1 isfaculative.
(USER:) why didn't you considerstreptococcus as a possibility for organism-1?
(TEIRESIAS:) The following rule could have beenused to determine that the identity of organism-1was streptococcus: RULE33.
9expert.doc EE 4524 Page 30
But clause 2 ("the morphology of the organism iscoccus") was already known to be false fororganism-1, so the rule was never tried.
9expert.doc EE 4524 Page 31
How does MYCIN know whatquestions to ask?
At each point, the question is determined byMYCIN's current hypothesis (andanswers to previous questions)
MYCIN is a backward-chaining system:Eg., to determine the cause of thepatient's illness, MYCIN looks for ruleswhich have a THEN clause suggestingdiseases;MYCIN then uses the IF clauses to setup subgoals, and looks for THENclauses of other rules to satisfy thesesubgoals, etc.
This approach makes it easier for thephysician to follow the "thought"process, and it simplifies the English-language interface
9expert.doc EE 4524 Page 32
MYCIN summary
... recommends therapies forpatients with bacterial infections
... uses IF-THEN rules (withcertainty factors) to representknowledge
... interacts with a physician toacquire clinical data
... asks questions based on currenthypothesis and known data
... reasons backward from its goalof recommending a therapy fora particular patient
... stores approx. 500 IF-THENrules, and can recognize about100 causes of bacterial infection
9expert.doc EE 4524 Page 33
TEIRESIAS summary
... serves as a front-end to MYCIN
... was the first program to provideexplanations of how conclusionswere reached
... intercepts questions such as "why"and "how" from the physician(i.e., why does MYCIN wantcertain information, and how didMYCIN reach a certainconclusion)
... TEIRESIAS can answer "why"questions by examining its internaltree of subgoals
... TEIRESIAS can answer "how"questions by identifying the piecesof evidence that supportedMYCIN's IF clauses
9expert.doc EE 4524 Page 34
Expert system shells
After MYCIN was built, someoneobserved that the knowledge basecould be replaced by completely newrules
MYCIN without its knowledge base wascalled EMYCIN (Empty MYCIN)(and was used to implement PUFF)
Today you can buy similar "shells" thatcontain a user interface, a reasoningsubsystem, and an explanationsubsystem
With such a shell, the user canconcentrate on the knowledge base
9expert.doc EE 4524 Page 35
In many expert systems, the rulesare written as follows:
symptom ⇒ disease
(the diagnosis must work from symptoms to find thecause)
But in reality, we know that
disease ⇒ symptom
Abductive reasoning is not truth-preserving:
P ⇒ QQ_________∴P
9expert.doc EE 4524 Page 36
Reasoning under uncertainty(Inexact reasoning)
We can attach "confidence" or"belief" values to
● the inference itself:A ⇒ B (with confidence 0.8)
● the evidence:A (which has confidence 0.6)
⇒ B
● both
9expert.doc EE 4524 Page 37
Our first impulse for inexact reasoning:use probability theory!
What is Pr(measles | spots) ?
Recall Bayes' theorem:
Pr( )Pr( ) Pr( )
Pr( )measles spots
spots measles measles
spots|
|=
Looks fine. Now we'd like toconsider other possiblediseases:
Pr( )Pr( ) Pr( )
Pr( )H spots
spots H H
spotsii i|
|=
If the diseases are exhaustive andmutually exclusive:
=∑
Pr( ) Pr( )
Pr( ) Pr( )
spots H H
spots H Hi i
i ii
|
|
9expert.doc EE 4524 Page 38
Now consider two differentsymptoms for one disease:
Pr( )Pr( ) Pr( )
Pr( )H spots fever
spots fever H H
spots feveri
i i||
∧ =∧
∧
Problem: how do we compute these ?Pr( )spots fever∧
Pr( )spots fever Hi∧ |
It is common (and absurd!) toassume that spots and fever areindependent:
Pr( ) Pr( ) Pr( )spots fever spots fever∧ =
To really use Bayes' theorem, we wouldneed probabilities for all possible
9expert.doc EE 4524 Page 39
combinations of symptoms in allconditional expressions: not feasible!
9expert.doc EE 4524 Page 40
Standard reasons why Bayesianreasoning cannot work:
● in "pure form" it requires animpossible number of probabilities
● the usual remedy is to imposeabsurd assumptions ofindependence
● knowing any probability may beunrealistic(usually just use statistical frequency)
● it only works for the single-diseasesituation
Still, it's a good starting point . . .
9expert.doc EE 4524 Page 41
MYCIN's Confidence Factors
a MYCIN rule: E ⇒ H (CF=x)
Confidence Factor:1.0 true with completeconfidence
-1.0 false with complete confidence
If x = 1.0 and E is a predicate, then wehave normal logic
CF(H|E) = MB(H|E) - MD(H|E)
MB: "measure of belief"MD: "measure of disbelief"
Each is in range [0, 1]When one is nonzero, the other is
normally zero
9expert.doc EE 4524 Page 42
Consider E E H1 2∧ ⇒ (CF = x)
If the Ei are all certain, then H has CF =x
If the Ei are not all certain, then weneed to "fold together" theconfidence factors
For conjunctive evidence:MB E E( )1 2∧ =
min( ( ), ( ))MB E MB E1 2MD E E( )1 2∧ =
max( ( ), ( ))MD E MD E1 2
Now consider E E H1 2∨ ⇒ (CF = x):For disjunctive evidence:MB E E( )1 2∨ =
max( ( ), ( ))MB E MB E1 2MD E E( )1 2∨ =
min( ( ), ( ))MD H MD H1 2
9expert.doc EE 4524 Page 43
What CF do we assign to H, foruncertain evidence P?
P H⇒ (CF = x)
MB H MB H CF P( ) '( ) ( , ( ))= max 0
MD H MD H CF P( ) '( ) ( , ( ))= max 0
9expert.doc EE 4524 Page 44
Now consider this:Rule 1: E H1 ⇒ (CF=x)Rule 2: E H2 ⇒ (CF=y)
If both Ei are true,then both should contribute tothe confidence that H is true:
MB H E E( )| 1 2∧ =0 11 2
1 2
1 2
MD H E E
MB H E MB H E otherwise
MB H E MB H E
( )
( ) ( )
( ) ( )
|
| |
| |
∧ =+
−
MD H E E( )| 1 2∧ =0 11 2
1 2
1 2
MB H E E
MD H E MD H E otherwise
MD H E MD H E
( )
( ) ( )
( ) ( )
|
| |
| |
∧ =+
−
(see text, p. 234)
9expert.doc EE 4524 Page 45
In MYCIN, rules are invoked bybackwards-chaining usingexhaustive depth-first search
Eg., find all rules that conclude theidentity of an organism
Eg., see if all conditions are met; ifnot, set up subgoals (based onIF clauses)
If -0.2 < CF < 0.2, the CF value isregarded as unknown. In thiscase, MYCIN asks the user.
9expert.doc EE 4524 Page 46
a different approach . . .
PLANT/ds
an expert system for diagnosingsoybean diseases
Rule form: extended propositionallogic
9expert.doc EE 4524 Page 47
PLANT/ds rules
Let x1, x2, ..., xn representdifferent “features” that can beobserved or measured
[x2 != 3] [x3 = 1, 3] v [x4 < 4]=> [decision = A]
(each [...] is called a “selector”; the first oneis TRUE if x2 is not equal to 3)
0.9 ([x1 = 3] [x3 >= 2])+ 0.1 ( [x3 = 2..4])
=> [decision = B](90% of the support comes from 2
selectors, and 10% from another)
9expert.doc EE 4524 Page 48
A sample PLANT/ds rule:
0.8 * ([time = Aug..Sept][precip = 0.8][fruiting bodies = present][stem cankers = above 2nd node][fruit pods = absent] )
+0.2 * ([temp >= normal]
[canker lesion color = brown][# years crop repeated = 2 yrs.])
=> [diagnosis = diaporthe stem canker]
9expert.doc EE 4524 Page 49
Rule evaluation:● each selector [var = value]
contributes a strength of evidencefrom 0 to 1
● if selector is not matched exactly,and is for a variable with linearlyordered domain, then a normaldistribution is assumed
● for conjunctions, evidence iscombined by product, min, oraverage
● for disjunctions, evidence iscombined by max or (w1+w2-w1w2)
● evidence weights are scaled bymodule coefficients and added to getstrength of evidence of conclusion
● The rule that scores best wins;Rules coming within 25% of best areconsidered alternativesRules below 0.8 are not reported
9expert.doc EE 4524 Page 50
PLANT/ds sessionEach question which follows will have a list of answers.
Choose the best answer, type the correspondingletter, then press return.
TIME OF OCCURRENCE?A. AprilB. MayC. JuneD. JulyE. AugustF. SeptemberG. OctoberH. Do not know** B
CONDITION OF LEAVES?A. NormalB. AbnormalC. Do not know** B
CONDITION OF STEM?A. NormalB. AbnormalC. Do not know** why
9expert.doc EE 4524 Page 51
IF[CONDITION OF STEM = NORMAL] then thefollowing are indicated:Downy MildewBacterial PustulePhyllosticta Leaf SpotAlternaria Leaf Spot
IF[CONDITION OF STEM = ABNORMAL] then thefollowing are indicated:Charcoal RotRhizoctonia Root RotPhytophthora RotBrown Stem RotAnthracnoseFrog Eye Leaf Spot
CONDITION OF STEM?A. NormalB. AbnormalC. Do not know** A
9expert.doc EE 4524 Page 52
CONDITION OF FRUIT-PODS?A. NormalB. DiseasedC. Few or none presentD. Does not applyE. Do not know** A
CONDITION OF SEED?A. NormalB. AbnormalC. Do not know** A
CONDITION OF ROOTS?A. NormalB. AbnormalC. Galls or cysts presentD. Do not know** A...
9expert.doc EE 4524 Page 53
CROPPING HISTORY?A. Crop different than last yearB. Crop same as last yearC. Crop same for last three yearsD. Crop same for four or more yearsE. Do not know** A
THE EVIDENCE PRESENTED SUGGESTS
BROWN SPOT WITH A DEGREE OF CONFIDENCE1.00
ALTERNATIVELY
PHYLLOSTICTA LEAF SPOT WITH A DEGREE OFCONFIDENCE 0.82
9expert.doc EE 4524 Page 54
The burden of theKnowledge Engineer
skill system rulesdiagnosing
soybeandiseases
PLANT/ds 25
identifyingbacteria
MYCIN 400
finding structureof organiccompounds
DENDRAL 445
playinggrandmasterchess
human 30,000
processing avisual scene
human ???
9expert.doc EE 4524 Page 55
So many rules!
Human experts are not very goodat writing rules
What if the computer could learnits own rules?!
9expert.doc EE 4524 Page 56
This was tried with PLANT/ds(Michalski & Chilausky, 1981,
Illinois)
Rules for diagnosing soybeandiseases were generated fromexamples that were correctlyclassified by disease type by ahuman expert
Surprise! (not really)
Machine-derived rules performedbetter than the rules given bythe human expert
original human rules 83% correctimproved human rules 93%machine-derived rules 99%
9expert.doc EE 4524 Page 57
How did the machine "learn" thecorrect rules?
data was collected for 350 sick plantsthought to suffer from one of 15diseases
the plant expert characterized eachdiseased plant using 35 differentfeatures(each plant was represented as a pointin a 35-dimensional space)
the plant expert divided the 350 data pointsinto 15 different classes (one class perdisease)
an inductive learning program generalizedfrom the given points to find simple rulesto describe each class(this is called learning from examples)
9expert.doc EE 4524 Page 58
the 15 rules (one rule for eachclass) were put into theknowledge base
these rules were tested using newcases of diseased plants
We could say that the machine"acquired knowledge" byexamining the given examples
The system "learns" the necessaryrules by performing inductiveinference (generalization) oversets of examples
Machine learning is important forbuilding large-scale expertsystems
9expert.doc EE 4524 Page 59
Knowledge-based systems:summary
Knowledge-based systems areways to capture and use theknowledge of human experts
Knowledge-based systems need aknowledge base and areasoning mechanism
IF-THEN rules are common, butother knowledge-representations are possible(eg., semantic nets)
Machine learning methods canhelp with large knowledgebases
More commercial successes herethan any other part of AI
9expert.doc EE 4524 Page 60
Knowledge-based systems:limitations
Knowledge-base generation andmaintenance are difficult chores
Knowledge-based systems "know"only the things in the knowledgebase
They do not know how their ruleswere developed
They do not know when to breaktheir own rules
They do not look at problems fromdifferent perspectives
Most cannot reason at multiplelevels
They typically cannot learn fromtheir own experiences