290 likes | 805 Views
Mycin. CIS 479/579 Bruce R. Maxim UM-Dearborn. Mycin. Task Prescribe drug regimen for infectious virus and diseases Rationale It was estimated that in 1972 physicians were over prescribing antibiotics by a factor of 10-20 times too much
E N D
Mycin CIS 479/579 Bruce R. Maxim UM-Dearborn
Mycin • Task • Prescribe drug regimen for infectious virus and diseases • Rationale • It was estimated that in 1972 physicians were over prescribing antibiotics by a factor of 10-20 times too much • Costly and breeds strains of bacteria resistant to treatment
Design Parameter • Program must be competent & easy to use • Must handle a large, changing body of knowledge • Interact with human users • Must take time into account • Work with incomplete or uncertain information
System Components • Consultation system • Asks questions • Draws conclusions • Gives advice • Explanation system • Translates rule to English before display • Rule acquisition/modification system
Rule Grammar <rules. ::= <premise> <action> <premise> ::= ($AND <condition> … <condition>) <condition> ::= (<predicate> <context> <parameter> <value>) | ($OR <condition> … <condition>) <action> ::= <conclusion> | <instruction>
Lisp Rule • Premise ($and (same cntxt gram gramneg) (same cntxt morph rod) (same cntxt air aerobic)) • Action (conclude cntxt class enterobacteriacae tally 0.8)
English Rule IF strain of org is gramneg and morphology of org is rod and aerobicity of org is aerobic THEN there is strongly suggestive evidence (0.8) that the class of org is enterobacteriacae
Static Data Structures • Simple lists • enumerate organisms and sterile sites known to system • Knowledge tables • contain clinical parameters and their values under various circumstances • Classification system • clinical parameters according the contexts in which they apply
Dynamic Data Structures • Context tree • serves to organize information relating to a particular patient • used to structure clinical problem and relate contexts to one another • rules are related to the context tree (although the rules themselves are not organized into either a decision tree or inference network)
Working Memory • Knowledge is represented using AOV triples with certainty factors (attribute object value certainty) (IDENT ORG1 ENTEROBACTERIACAE 0.8)
Control Structure • Backward chaining • helps to keep it focuses • facilitates backward reasoning from top level goal for all queries
Inference Strategy • Subgoals are generalized (i.e. match with variables) when possible • All applicable rules are evaluated before reaching a decision • Facts with certainities between –0.2 and +0.2 are treated as unknown • Mycin asks for lab for some facts before attempting a deduction • A list of rules that fail under the current context is maintained to avoid re-evaluation • Premises are evaluated based on known fact before search is allowed
Goal Rule IF there is an org requiring therapy and consideration has been given to possibilty of other orgs requiring therapy THEN compile a list of possible therapies and select the best alternative from list
Consultation Procedure • Create a patient context as the top level node in the context tree • Attempt to apply the goal rule to this particular patient context • Context tree is fleshed out in an effort to accumulate evidence from user query or inference • Each node contains accumulated evidence including “lab data” to allow alternation between question selection and rule invocation
Mutually Recursive Procedures • Monitor • attempts to evaluate premise of current rule • if it fails rule is discarded and next rule from list is examined (restricted by context) • Findout • gathers evidence for and against rule premise • if question can be asked control returns to Monitor with answer • if no question new list of rules to determine truth of rule premise is returned to Monitor
Mycin Rules • Had 200 rules in 1976 • Meta-rules • rule pruners similar to alpha/beta cutoffs • rules to reorder relevant domain rules • general (domain free) problem solving heuristics • some forward (antecedent) reasoning to cut stupid questions (i.e. skip pregnancy questions for males)
Explanation System • Can display rule being invoked at any point in consultation • Record rule invocation and associates them with questions asked and rules invoked • Use rule index to retrieve particular rules in answer to questions • Why and how questions answered using goal tree
Rule Acquisition System • Domain experts allowed to enter and change rules • Rules translated to Lisp and rule numbers added to “Look-ahead” and “Updated-by” lists • Does not catch contractions and inconsistencies in large rule-bases
Evaluation 1974 • Panel of 5 experts approve 72% of Mycin’s recommendations for 15 patients 1976 • 8 experts (5 faculty, 1 resident, 1 med student, 1 research fellow) made drug recommendations for 10 patients • Mycin had best match (52%) with actual drug recommendations used by attending physician
Summary • Mycin combines the advantages of general rule-based system with the advantages of an “inexact” reasoning system • Mycin has not addressed • how to convert from human terms to certainties • how to normalize across different people’s • how far to propagate certainty factor changes based on new evidence • how to provide feedback to database to improve certainty factor accuracy
Emycin • One of the first expert systems shells • Basically Mycin with its domain specific knowledge removed • Uses backward chaining with Mycin’s antecedent rule processing • Includes some tools to allow KE’s to debug large systems
Monitoring Behavior • Explain • essentially a trace of rules involved to make inferences • Test • allows comparison of current run with stored run and prints list of discprepencies • Review • allows expert to examine system conclusions on stored cases
Sacon • Emycin had an English-like rule language called Sacon IF composition = list of methods and error < 5 and nd-cycles > 0.5 and THEN ss-stree = fatique
Emycin Weaknesses • Emycin approach is only useful in domain where task is to provide consultative advice • Use of certainty factors smacks of probability theory without having assurance of independence • Certainty factors may not be valid in every domain • Approach was derived from Mycin which was implemented to be very domain specific
Teiresias • Created to do knowledge acquisition and truth maintenance for Mycin • Designed to identify • gaps in rule set (i.e. combinations of conditions not covered) • overlapping rules that generate inconsistency and redundancy • rules that become obsolete due to new discoveries
Teiresias • System was able to use syntactic and semantic knowledge from domain to assign blame to particular rules and make suggestions to problem solution • Allows expert to focus on program errors like drawing false conclusions or asking irrelevant questions • Program can be stopped on errors and make use of Mycin’s how/why facilities
Neomycin • Uses domain independent strategy for doing diagnosis based on a psychological model of diagnostic behavior • Uses forward chaining from known data and generates new hypotheses • Maintains working memory of alternate hypotheses explored using a “group and differentiate” strategy
Rule-based Systems • Can be used a a foundation for intelligent systems work since it makes use of a homogeneous representation of knowledge • Allow incremental knowledge growth by adding more if…then type rules • Allow unplanned, but useful rule interactions