1. © C. Kemke 1Classification Problem Solving COMP 4200: Expert Systems Dr. Christel Kemke Department of Computer Science University of Manitoba

2. © C. Kemke 2Classification Problem Solving  Heuristic Classification  MYCIN  MUD cf. Jackson, Chapters 11 and 12

3. © C. Kemke 3Classification Problem Solving Solution is selected from pre-defined set  example: diagnosis of one of several described illnesses  classification must not be direct and one-to-one but may involve intermediate levels of conclusions and uncertainty or ambiguity  heuristic classification based on such "inexact" rules

4. © C. Kemke 4Classification Problem Solving Heuristic Classification  Hierarchical organization of classes  sub-classes inherit discriminative features  sub-classes mutually exclusive  solution classes can be enumerated  hierarchy can involve different types of classifications  use data to find suitable classes in hierarchy - heuristic classification - up to final diagnosis / solution class, in a stepwise hierarchical classification process

5. © C. Kemke 5Classification Problem Solving Classification Problem Solving Steps  Data Abstraction  map raw data into relevant categories, like 'blood sugar critical' or 'temperature very high' instead of dealing with exact measured values  Heuristic Match  hypothesize broad class of solutions, e.g. for diseases, the symptom 'high body temperature' or 'fever' indicates an infection.  Solution Refinement  identify and rank solutions within broad solution class, e.g. by further conclusions based on given data or through collecting new data.

6. © C. Kemke 6Classification Problem Solving (Data) Abstraction  Definitional  essential features of class / category  e.g. mammals have life-born; animals have an excitable cell membrane  Qualitative  abstracting over quantitative measures  e.g. high/medium/low cholesterol; battery status okay/low/empty  Generalization  abstracting in a hierarchy  Leukopenia is a kind of immunosuppresion;...

7. © C. Kemke 7Classification Problem Solving Classification Problem Solving: Mycin Compromised Host Immunosuppressed Host Leukopenia Low WBC WBC < 2.5 Gram-negative Infection E. Coli Infection D a t a A b st r a ct i o n Classification Diagnosis RefinementRefinement

8. © C. Kemke 8Classification Problem Solving Generic XPS Task (Chandrasekaran)  Domain Knowledge Description of various forms of domain knowledge and structure, organization of this knowledge  Control Regimes Control flow in Problem Solving / XPS Performance

9. © C. Kemke 9Classification Problem Solving Tasks 1 (Chandrasekaran) Viewpoint of Hypothesis Evaluation  Hierarchical Classification select explanatory hypothesis from hierarchically organized space of alternatives; refine hypothesis to account for data; test against data  Hypothesis Matching weigh evidence regarding goodness of fit of hypothesis with respect to observed facts (data); include prior probability of hypothesis, match with data, comparison of competing hypothesis

10. © C. Kemke 10Classification Problem Solving Tasks 2 (Chandrasekaran)  Knowledge-directed Information Passing Background knowledge and rules about relations (causal, inferential) in the domain, e.g. if it's below -40C a car may freeze and thus not start.

11. © C. Kemke 11Classification Problem Solving MUD

12. © C. Kemke 12Classification Problem Solving Classification Problem Solving II MUD Expert System  MUD: determine cause of observed symptoms in drilling fluid domain  MORE: classification involves causal model of the domain: causal chain from factors in the environment (diagnosis) to observable properties of the drilling fluid (symptoms)  rules involve evidence factors for H being true if Condition is present / not present

13. © C. Kemke 13Classification Problem Solving MUD – Part of Causal Model shale contamination increase in solids water influx decrease in density increase in viscosity increase in unemulsified water gradual rapid oil mud in use MBT

14. © C. Kemke 14Classification Problem Solving MUD – Knowledge  Symptoms – observed during diagnosis, search for explanation, e.g. increase in viscosity  Attributes – further discriminate symptoms, e.g. rapid decrease in attribute value  Events – possible causes of symptoms; are hypotheses, e.g. water influx  Background Conditions – influence a-priori probabilities of events, causes, and conditional probabilities of effects (e.g. oil mud in use)  Tests to determine background conditions (MBT to test increase of solids in drilling fluid)  Test Conditions which influence accuracy of tests

15. © C. Kemke 15Classification Problem Solving MUD – Knowledge Acquisition Different types of knowledge acquired  Differentiation – seek symptoms S that distinguish between hypotheses (diagnosis) H  Frequency Conditionalization – determine a-priori probabilities of H involving background conditions  Symptom Distinction – identify special properties of symptoms that indicate underlying cause  Symptom Conditionalization – find conditions which indicate that diagnosis H is true if symptoms S are present

16. © C. Kemke 16Classification Problem Solving MUD – Knowledge Acquisition  Path Division – uncover intermediate events I between symptoms and hypotheses which have stronger (statistical) connection to H than S, i.e. I is better indication for H than S alone.  Path Differentiation – uncover intermediate events between symptoms and hypotheses in order to distinguish different H with similar S.  Test Differentiation – determine the degree of confidence in test results (reliability of tests).  Test Conditionalization – determine background conditions which influence reliability of tests.

17. © C. Kemke 17Classification Problem Solving Probability Theory Background Probability P of a hypothesis H:  unconditional (a-priori, prior) probability probability of hypothesis H on it’s own: P(H) example: throwing a six with a perfect dice  conditional (posterior) probability probability of hypothesis H depending on observed event E: P(H|E)  example: H  outside temperature is below 0  C; event E  the date

18. © C. Kemke 18Classification Problem Solving Probability Theory Example 1 Unconditional Probability Determine the probabilities P of the hypotheses H: example: throwing a perfect dice 1.What is P(H1) with H1=dice shows a 6? 2.What is P(H2) with H2=dice shows a 1? 3.What is P(H) with H=dice shows a 6 or a 1? 4.What is P(H1 or H2)?

19. © C. Kemke 19Classification Problem Solving Probability Theory Example 2 Conditional Probability Guess the probability P of the hypothesis H: example: Is it below 0°C outside? Hypothesis H  outside temperature is below 0°C; Event E  the current date 1.What could P(H|E) be? 2.What would then be the probability that it is equal or above 0°C outside, i.e. P(not(H)|E) ? Note: H and not(H) are complementary hypotheses.