Expert System in Artificial Intelligence
An Expert System is a computer system that simulates the decision making process of a human expert in a specific domain.
Human expert has two kinds of knowledge:
– public knowledge: published facts, definitions and theories
– private knowledge: heuristics, or rules of thumb
Image analysis, Medical Diagnosis, Air Traffic Control, Budget Planning, Robotics, Chemical Structure, Crop Estimates and so on.
History of Expert System:
– 1966 DENDRAL, molecular structure
– 1971 MYCIN, blood diseases
– 1980 XCON, configuring computer ordering
Characteristics of an Expert System:
a. Require symbolic representation
b. Numerous alternatives at each step
c. Many combinations of alternatives need to be considered
d. Important decisions are scalar not binary
– permanent, fast processing, consistent, quick replication, affordable
– narrow focus, lacks creative ability, instruction machine knowledge
– knowledge implemented as a RULE
– a rule is an IF-THEN type statement
So, an expert system is called ruled-based systems or production systems
Japan Bullet Train requires 100 rules.
– Rule Base, data structure containing rules
– Working Memory, data structure storing information about a specific run
– Inference Engine, procedure for matching current knowledge with the rules to produce new knowledge.
a. from ground up for each problem program inference engine, knowledge base and working memory.
b. program general IE, KB and WM and add required knowledge.
c. buy a general IE and KB, and add specific knowledge.
e.g EMYCIN, OPS5, EXSYS
– it is the process of filling in an expert system knowledge base
Expert System <——–Knowledge Engineer—————>Domain Expert
Knowledge engineer queries problems and get answers and solution from domain expert, and fill the expert system with domain rules, rules of thumb and strategies.
Data implies numbers.
Information is interpreted, categorized and applied
Knowledge is extracted required information.
Types of knowledge:
a. Procedural – knowing how
b. Declarative – knowing what
c. Semantic – cognitive structure & symbolism
d. Episodic – experiential & temporal
e. Meta – knowledge about knowledge
a. Rule based- IF THEN
b. Object based – Frames, scripts, semantic networks
c. Example based – case based reasoning
knowledge Base size:
– expertise in a profession requires 10,000 rules
– limit of human expertise 100, 000 rules
– expert level of competence in narrow area 500 -1000 rules
– commercially practical solution requires 50 rules
– interesting demonstration of technology requires 50 rules
– convincing demonstration of a knowledge system requires 250 rules
a. Relationship (FACT) – if battery dead, then car will not start
b. Recommendation – if car do not start, then take a cab
c. Directive – if car not start, fuel ok, then check electrical system
d. Heuristic – if car not start and is a 1957 ford, then check float.
Consists of 3 steps:
1. Match: rules are compared with working memory
2. Conflict Resolution: select or enable a single rule
3. Execute: fire selected rule
– Rule schedule, Rule Activation, uncertainty management, how processor, why processor
The goal of rule schedule is to use current knowledge to determine which rule or rules to activate in the next reasoning cycle. Two approaches:
Forward chaining (data driven)
Backward chaining (goal directed)
– Semantic Networks
– O-A-V Triplets
Uncertain Information is deficient in one or more of the following ways: partial, not reliable, conflicting, imprecise.
It is techniques used to effectively handle uncertainty. It is tolerant to imprecision, uncertainty and partial truth. Its techniques are probability theory, utility theory and decision theory, certainty factors, Dempster-Schafer Calculus, Fuzzy Logic and Genetic algorithm.
Certainty Factors Calculus:
– Range from -100 to +100, +20 and above is considered true in MYCIN.
Each proposition A has initial Certainty Measure (CM) = 0.
If A -> B with CF = x, and A is true then new CM for B is x.
– If Then rule, multiple uncertainties, uncertain rules ct = crule * cpremise/100, accumulating positive evidence ct = cf1 + cf2 % (100 – cf1), accumulating negative evidence ct = -(|cf1| + |cf2| % (100 – |cf1|)) and mixed evidence cft = 100(cf1 + cf2)/(100-A), A = min(|cf1, |cf2|).
Management of Uncertainty:
1. Representation of uncertain information
2. Combination of bodies of uncertain information
3. Drawing inference using uncertain information
– mathematical theory of evidence
– B(A), P*(A) = 1 – Doubt(A), Doubt(A) = B(-A)
A fuzzy set is a collection of objects with imprecise membership.
Fuzzy inference involves three steps:
1. Fuzzification: create fuzzy representation of the conditions
2. Inference – determine the fuzzy representation of the conclusion.
3. Defuzzification – convert the conclusion to a hard result.
Bayes Inference Network:
A set of rules, which does not contain cyclic rules, determine a DAG, directed acyclic graph
It is a directed random search procedure.
– random population of solutions, evaluate, random selection, random mating, random mutation, new population.
Components: Populations and Fitness function.