Lecture 9 - Expert Systems

• the Classical AI - what most people think of when mention AI
• Turing Test is too hard - need to reduce to limited microworlds
• GPS (Newell+Simon 1963)
• could prove theorems
• solve recreational logic problems (backwards chaining)
• The Logic Theorist
• proved much of Principia Mathematica
• showed utility of AI techniques
• but this approach failed on real world problems (unstructured, ill-defined)
• could not translate problem into appropriate form
• could not prune vast search space
• how do people cope?
• use semilogical methods
• use of logic as aid and rationalization
• but intuition as main engine
• approximate matching of many previously seen patterns to appropriate action
• high IQ machine is not enough -experts have expert knowledge (Dendral)
• Expert system architecture
• Knowledge Base (KB)
• Inference engine
• user interface
• KB acquired by interviewing experts or from examples
• knowledge engineers
• unlike system analysts they don’t try to create new algorithm
• however similarities do exist
• choose logic - choose programming language
• setup inference engine - compile
• acquire knowledge - code program
• infer new facts - run program
• stages
1. acquire domain knowledge
2. decide on variables, predicates, rules, etc.
3. code general knowledge
4. encode specific case
5. query about specific case and check answers
• inference engine
• may be forward chaining, backward chaining, probabilistic, etc.
• treatment of unknown (missing) or uncertain data (KB updatable by user ?)
• need more than Prolog (need HMI, trace, dialog memory, chaining plan, etc)
• but usually written in Prolog/LISP/special ES shell
1. each fact/rule is independent of others (easy to debug, monotonicity)
2. no hard encoding of rules - easy to revise facts, rules, even flow
• user interface needs to be friendly, and has to ...
1. aid in knowledge acquisition
2. explain its reasoning and conclusions
3. interact with user to get problem specific data
4. direct user to acquire more data when needed
• easiest to understand by concrete examples
• DENDRAL
• MACSYMA
• MYCIN
• R1(XCON)
• INTERNIST
• PROSPECTOR
• HEARSAY
• simple example - animal world