1. Expert Systems
expert system (or knowledge-based system): a program which encapsulates knowledge from some domain, normally obtained from a human expert in that domain
components:
Knowledge base (KB): repository of rules, facts (productions)
working memory: (if forward chaining used)
inference engine: the deduction system used to infer results from user input and KB
user interface: interfaces with user
external control + monitoring: access external databases, control aircraft rudders,...

2. Expert Systems Why use expert systems:
commercial viability: whereas there may be only a few experts whose time is expensive and rare, you can have many expert systems
expert systems can be used anywhere, anytime
expert systems can have more knowledge than experts
expert systems can explain their line of reasoning
commercially beneficial: the first commercial product of AI
Weaknesses:
expert systems are as sound as their KB; errors in rules mean errors in diagnoses
automatic error correction, learning is difficult (although machine learning research may change this)
little common sense reasoning: “idiot savantes”

3. Steps in expert system building
expert system building is a software engineering task, but with a twist: knowledge base
KB will be continually refined, corrected, updated
expert system design & implementation steps
1. initial proposal - identification of problem, expert(s), benefits
2. create a prototype
3. knowledge engineering: interview expert for many weeks (months)
4. implement expert’s expertise in KB
5. test expert system; go to 3 (let expert see results)
6. put in production ; refine errors when found

4. Knowledge engineering
Knowledge engineering: the extraction of knowledge from an expert, and encoding it into machine-inferrable form
K.E. is the most difficult part of expert system implementation
the knowledge engineer must extract detailed info from an expert; but s/he is not expected to understand it
the expert relies on years of experience and intuition: asking them to deconstruct knowledge is difficult and frustrating
can also be political factors
many strategies for knowledge engineering
multiple KE’s
multiple experts
watch expert in the field
continual expert feedback on expert system performance

5. Explanation
Recall AND/OR trees: when you are asking why a particular action was performed at a node in the tree...
how questions: recite the children
why questions: recite the parent
repeated how or why: go down/up further in the tree
thus backward chaining expert systems naturally support explanation
can be done during inference: when user is asked for data, s/he can ask “why” and a reason is given
after expert advice is given, user can ask “how”, “why”, and even “why not” (ie. why did some rule fail)
a great tool for KB debugging!
forward chaining: explanation not as structured; usually, rule ID’s are generated showing which rules fired; canned text possible too

6. Reactive systems
Forward- and backward-chaining expert systems can be reactive: explicit actions can be carried out (in addition to, or instead of, ‘advice’ or diagnoses)
most common reactions: add a new fact, delete an existing fact from working memory
with forward chaining: since assertions in working memory will fire rules, you may need to delete some in order to stop a rule from firing, and thus moving the deduction forward
one potential danger of forward-chaining systems: unstructured KB and inference
dump a load of rules and facts into KB, working memory
let the forward chaining execute until a solution obtained
but this permits ad hoc expert system design

7. Probabilistic expert systems
Expertise is often intuitive, and facts can have ‘degrees’ of truth
much research has been devoted to inventing new models of probabilistic deduction: eg. Certainty factors
fuzzy logic
statistical / Bayesian deduction
multi-valued logics: true, false, unknown
typical rule: (eg). IF patient has a fever (CF 0.90)
patient’s forehead is blue (CF .75)
THEN illness is Cat Scratch Fever (CF .5)
weakness: these certainty factors are not statistically sound values; rather “pulled out of the air” by expert or KE
CF’s are combined during deduction; multiple results can arise
how legitimate are they? “GIGO”

8. Other expert system paradigms
1. Neural nets & machine learning
neural net: network of linked nodes; threshold functions fire nodes
examples of desired behaviour are obtained from expert, and converted to numeric vectors for a neural network
the network is trained with an example set; output values are refined until network gives desired behaviour
final system accepts general questions (converted to numeric values) and output reflects appropriate diagnoses
rules can be extracted from neural net as well
strengths:
a ‘bottom-up’ approach to KB (vs. structured)
perhaps human expertise is similarly acquired?
weaknesses:
what the heck is going on in the net???

9. Other expert systems paradigms
2. Inductive inference & machine learning
rather than extract rules explicitly from expert, it’s easier to get table of example behaviours (like with neural nets)
inductive inference: generalizes from finite example set to whole domain
using ideas from information theory, a refined decision tree (if-then-else) is derived from examples
strengths:
examples are intuitive (simplifies knowledge acquisition phase))
can often find interesting patterns unknown by experts
generates ‘traditional’ rule-based KB’s
weaknesses:
example set must be representative
noise can be bad

10. Examples expert systems
1. Medical expert systems
medical expert systems are an active area
humanitarian uses: remote areas can have world-class expertise, eg. a laptop and CD ROM
medical science is complex; expert systems are intelligent advisors and assistants
MYCIN - diagnoses bacterial blood infections
uses certainty factors: ranked diagnoses are generated
EMYCIN: empty MYCIN - the “shell” of MYCIN for use in other applications
CADUCEUS:
expert system with entire internal medicine KB: huge!
it’s been active since 1970’s

11. Example expert systems
2. XCON
configures VAX systems from user requirements
DEC uses it; has increased throughput of system design
10,000 rules
said to have replaced X system designers with 4X expert system maintainers!
3. DELTA/CATS
diesel electric locomotive trouble shooter
problem: system maintenance
4. Jonathan’s Wave
commodities trading
incorporates several expert’s approaches
runs in Prolog and C
lots of AI groups on Wall Street!

12. Example expert systems
5. Codecheck
expert system to evaluate C source code
complexity, formatting, standards and portability adherence
identifies overly complex code (prime source of program problems!)
6. CHORAL
expert system for harmonizing chorales in the style of J.S. Bach
similar system: harmonic analysis of tonal music
rules obtained from: studying composers music, musicologists
some systems used for composition (with interesting results)

13. The state of the art Expert systems are now ubiquitous
still lots of active research:
probabilistic knowledge and deduction
large system maintenance
knowledge acquisition
popular domains
industrial, financial, medical
legal: interpret laws “by the book”
education: intelligent tutoring
Some domains remain difficult to implement
those requiring ‘artistic’ creativity and skill
those requiring common sense: the idiot savante syndrome