EXPERT SYSTEMS AND KNOWLEDGE REPRESENTATION Ivan Bratko Faculty of Computer and Info. Sc. University of Ljubljana

EPERT SYSTEMS An expert system behaves like an expert in some narrow area of expertise Typical examples: Diagnosis in an area of medicine Adjusting between economy class and business class seats for future flights Diagnosing paper problems in rotary printing Very fashionable area of AI in period 1985-95

SOME FAMOUS EXPERT SYSTEMS MYCIN (Shortliffe, Feigenbaum, late seventies) Diagnosis of infections PROSPECTOR (Buchanan, et al. ~1980) Ore and oil exploration INTERNIST (Popple, mid eighties) Internal medicine

Structure of expert systems

FEATURES OF EXPERT SYSTEMS Problem solving in the area of expertise Interaction with user during and after problem solving Relying heavily on domain knowledge Explanation: ability of explain results to user

REPRESENTING KNOWLEDGE WITH IF-THEN RULES Traditionally the most popular form of knowledge representation in expert systems Examples of rules: if condition P then conclusion C if situation S then action A if conditions C1 and C2 hold then condition C does not hold

DESIRABLE FEATURES OF RULES Modularity: each rule defines a relatively independent piece of knowledge Incrementability: new rules added (relatively independently) of other rules Support explanation Can represent uncertainty

TYPICAL TYPES OF EXPLANATION How explanation Answers user’s questions of form: How did you reach this conclusion? Why explanation Answers users questions: Why do you need this information?

RULES CAN ALSO HANDLE UNCERTAINTY If condition A then conclusion B with certainty F

EXAMPLE RULE FROM MYCIN if 1 the infection is primary bacteremia, and 2 the site of the culture is one of the sterilesites, and 3 the suspected portal of entry of the organism is the gastrointestinal tract then there is suggestive evidence (0.7) that the identity of the organism is bacteroides.

EXAMPLE RULES FROM AL/X Diagnosing equipment failure on oil platforms, Reiter 1980 if the pressure in V-01 reached relief valve lift pressure then the relief valve on V-01 has lifted [N = 0.005, S = 400] NOT the pressure in V-01 reached relief valve lift pressure, and the relief valve on V-01 has lifted the V-01 relief valve opened early (the set pressure has drifted) [N = 0.001, S = 2000]

EXAMPLE RULE FROM AL3 Game playing, Bratko 1982 if 1 there is a hypothesis, H, that a plan P succeeds, and 2 there are two hypotheses, H1, that a plan R1 refutes plan P, and H2, that a plan R2 refutes plan P, and 3 there are facts: H1 is false, and H2 is false then 1 generate the hypothesis, H3, that the combined plan ‘R1 or R2' refutes plan P, and 2 generate the fact: H3 implies not(H)

A TOY KNOWLEDGE BASE: WATER IN FLAT Flat floor plan Inference network

IMPLEMENTING THE ‘LEAKS’ EXPERT SYSTEM Possible directly as Prolog rules: leak_in_bathroom :- hall_wet, kitchen_dry. Employ Prolog interpreter as inference engine Deficiences of this approach: Limited syntax Limited inference (Prolog’s backward chaining) Limited explanation

TAILOR THE SYNTAX WITH OPERATORS Rules: if hall_wet and kitchen_dry then leak_in_bathroom. Facts: fact( hall_wet). fact( bathroom_dry). fact( window_closed).

BACKWARD CHAINING RULE INTERPRETER is_true( P) :- fact( P). if Condition then P, % A relevant rule is_true( Condition). % whose condition is true is_true( P1 and P2) :- is_true( P1), is_true( P2). is_true( P1 or P2) :- is_true( P1) ; is_true( P2).

A FORWARD CHAINING RULE INTERPRETER new_derived_fact( P), % A new fact !, write( 'Derived: '), write( P), nl, assert( fact( P)), forward % Continue ; write( 'No more facts'). % All facts derived

FORWARD CHAINING INTERPRETER, CTD. new_derived_fact( Concl) :- if Cond then Concl, % A rule not fact( Concl), % Rule's conclusion not yet a fact composed_fact( Cond). % Condition true? composed_fact( Cond) :- fact( Cond). % Simple fact composed_fact( Cond1 and Cond2) :- composed_fact( Cond1), composed_fact( Cond2). % Both conjuncts true composed_fact( Cond1 or Cond2) :- composed_fact( Cond1) ; composed_fact( Cond2).

FORWARD VS. BACKWARD CHAINING Inference chains connect various types of info.: data ... goals evidence ... hypotheses findings, observations ... explanations, diagnoses manifestations ... diagnoses, causes Backward chaining: “goal driven” Forward chaining: “data driven”

FORWARD VS. BACKWARD CHAINING What is better? Checking a given hypothesis: backward more natural If there are many possible hypotheses: forward more natural (e.g. monitoring: data-driven) Often in expert reasoning: combination of both e.g. medical diagnosis water in flat: observe hall_wet, infer forward leak_in_bathroom, check backward kitchen_dry

GENERATING EXPLANATION How explanation: How did you find this answer? Explanation = proof tree of final conclusion E.g.: System: There is leak in kitchen. User: How did you get this?

% is_true( P, Proof) Proof is a proof that P is true :- op( 800, xfx, <=). is_true( P, P) :- fact( P). is_true( P, P <= CondProof) :- if Cond then P, is_true( Cond, CondProof). is_true( P1 and P2, Proof1 and Proof2) :- is_true( P1, Proof1), is_true( P2, Proof2). is_true( P1 or P2, Proof) :- is_true( P1, Proof) ; is_true( P2, Proof).

WHY EXPLANTION Exploring “leak in kitchen”, system asks: Was there rain? User (not sure) asks: Why do you need this? System: To explore whether water came from outside Why explanation = chain of rules between current goal and original (main) goal

UNCERTAINTY Propositions are qualified with: “certainty factors”, “degree of belief”, “subjective probability”, ... Proposition: CertaintyFactor if Condition then Conclusion: Certainty

A SIMPLE SCHEME FOR HANDLING UNCERTAINTY c( P1 and P2) = min( c(P1), c(P2)) c(P1 or P2) = max( c(P1), c(P2)) Rule “if P1 then P2: C” implies: c(P2) = c(P1) * C

DIFFICULTIES IN HANDLING UNCERTAINTY In our simple scheme, for example: Let c(a) = 0.5, c(b) = 0, then c( a or b) = 0.5 Now, suppose that c(b) increases to 0.5 Still: c( a or b) = 0.5 This is counter intuitive! Proper probability calculus would fix this Problem with probability calculus: Needs conditional probabilities - too many?!?

TWO SCHOOLS RE. UNCERTAINTY School 1: Probabilities impractical! Humans don’t think in terms of probability calculus anyway! School 2: Probabilities are mathematically sound and are the right way! Historical development of this debate: School 2 eventually prevailed - Bayesian nets, see e.g. Russell & Norvig 2003 (AI, Modern Approach), Bratko 2001 (Prolog Programming for AI)