An EXPERT SYSTEM is a computer program that emulates the DECISION-MAKING ABILITY of a human expert in a narrow domain, reasoning from a body of knowledge represented mainly as IF-THEN RULES. THE DEFINING CHARACTERISTICS: · expertise in a NARROW, well-bounded domain · knowledge held EXPLICITLY and SEPARATELY from the program logic · reasoning under UNCERTAINTY, with a confidence attached to conclusions · the ability to EXPLAIN its reasoning · performance at or near the level of a human specialist THE CENTRAL ARCHITECTURAL IDEA, and the one that made these systems possible: SEPARATE THE KNOWLEDGE FROM THE INFERENCE. A conventional program mixes domain knowledge into its control flow — the rules ARE the if-statements. Change a rule and you edit and recompile code. An expert system holds the rules as DATA, interpreted by a general INFERENCE ENGINE. A domain expert can then add a rule without a programmer, and the same engine serves any domain. THIS IS THE DECLARATIVE-OVER-PROCEDURAL ARGUMENT from the KR section, realised as a product architecture — and it is why "expert system shells" could be sold as generic tools. HOW AN EXPERT SYSTEM DIFFERS FROM A CONVENTIONAL PROGRAM: CONVENTIONAL PROGRAM EXPERT SYSTEM ───────────────────────────────────────────────────────────── knowledge embedded in code knowledge in a separate KB algorithmic — a fixed heuristic — searches for a procedure plausible conclusion processes data processes knowledge no explanation explains its reasoning right or wrong degrees of confidence changing logic means changing logic means editing recompiling a rule ───────────────────────────────────────────────────────────── THE HISTORICAL LANDMARKS, from the AI history topic: DENDRAL (1965–) inferring molecular structure from mass spectrometry — the first successful knowledge-intensive program, and the origin of the insight that DOMAIN KNOWLEDGE BEATS CLEVER SEARCH MYCIN (1972–) diagnosing bacterial blood infections, performing at specialist level with about 450 rules, and introducing CERTAINTY FACTORS for uncertainty PROSPECTOR mineral exploration; famously identified a molybdenum deposit XCON / R1 (1980) configuring VAX computer orders at Digital; grew to several thousand rules and reportedly saved tens of millions of dollars a year — the commercial proof

THE COMPONENTS of an expert system. This diagram and the component list are the guaranteed exam question. ┌──────────────────────────────────────────────────┐ │ USER │ └───────────────────────┬──────────────────────────┘ │ ┌───────────────────────▼──────────────────────────┐ │ USER INTERFACE │ │ (questions, answers, explanations) │ └───────┬──────────────────────────────┬───────────┘ │ │ ┌───────▼─────────┐ ┌─────────▼───────────┐ │ INFERENCE │◀────────▶│ EXPLANATION │ │ ENGINE │ │ FACILITY │ └───┬─────────┬───┘ └─────────────────────┘ │ │ ┌──────▼───┐ ┌──▼──────────────┐ │ KNOWLEDGE│ │ WORKING MEMORY │ │ BASE │ │ (the facts of │ │ (rules) │ │ this case) │ └────▲─────┘ └─────────────────┘ │ ┌────┴──────────────────┐ ┌──────────────────────┐ │ KNOWLEDGE ACQUISITION │◀────│ DOMAIN EXPERT + │ │ SUBSYSTEM │ │ KNOWLEDGE ENGINEER │ └───────────────────────┘ └──────────────────────┘ 1. KNOWLEDGE BASE The domain knowledge, usually as production rules: IF the organism is gram-positive AND the morphology is coccus AND the growth is in chains THEN the identity is streptococcus (CF 0.7) Also holds facts and, in frame-based systems, taxonomies. IT IS THE SYSTEM'S VALUE AND THE HARDEST PART TO BUILD. 2. WORKING MEMORY (also the fact base, or context) The facts known about THE CURRENT CASE — this patient, this order. Starts nearly empty, fills as the consultation proceeds, and is discarded afterwards. THE DISTINCTION FROM THE KNOWLEDGE BASE MATTERS: the KB is general and permanent, working memory is specific and transient. 3. INFERENCE ENGINE The general reasoning mechanism. It repeatedly: MATCH find rules whose premises are satisfied by working memory SELECT resolve conflicts when several rules match EXECUTE apply the chosen rule, adding its conclusion to working memory — the RECOGNISE-ACT CYCLE. It supports FORWARD and BACKWARD CHAINING (from the propositional logic topic) and handles the propagation of certainty. CONFLICT RESOLUTION STRATEGIES, when several rules match: specificity — prefer the rule with more premises recency — prefer the rule matching the newest fact priority — an explicit ordering supplied by the author refractoriness — do not fire the same rule on the same facts twice, which prevents infinite loops 4. EXPLANATION FACILITY Answers HOW and WHY: WHY are you asking me this? → shows the rule currently being pursued HOW did you conclude that? → shows the chain of rules that fired THIS IS NOT A COSMETIC FEATURE. A doctor will not act on a diagnosis they cannot check, so explanation was what made MYCIN acceptable to clinicians at all — and it is the property the AI applications topic identified as a regulatory requirement for opaque models today. 5. KNOWLEDGE ACQUISITION SUBSYSTEM Tools for entering and validating rules — an editor, a consistency checker, sometimes a rule-induction tool. Its inadequacy is the ACQUISITION BOTTLENECK that ended the boom, and it gets its own topic next. 6. USER INTERFACE Conducts the consultation, usually as a dialogue asking only for the facts the current line of reasoning needs. AN EXPERT SYSTEM SHELL is components 3, 4, 5 and 6 without a knowledge base — a generic engine into which any domain's rules can be loaded. EMYCIN was MYCIN with the medicine removed, and it is the origin of the whole commercial tools industry.

HANDLING UNCERTAINTY — MYCIN's CERTAINTY FACTOR approach, which predates the probabilistic methods of the previous section and is still what the syllabus expects. A CERTAINTY FACTOR CF ranges from −1 to +1: +1 known true 0 no evidence either way −1 known false It is defined from a MEASURE OF BELIEF MB and a MEASURE OF DISBELIEF MD: CF(h, e) = MB(h, e) − MD(h, e) PROPAGATING CERTAINTY THROUGH ONE RULE: CF(conclusion) = CF(premise) × CF(rule) WORKED: a premise established with CF 0.8, by a rule of strength 0.7: CF(conclusion) = 0.8 × 0.7 = 0.56 CONFIDENCE ONLY EVER DECREASES ALONG A CHAIN, which is intuitively right — conclusions drawn at three removes are less certain than their evidence. COMBINING TWO RULES THAT SUPPORT THE SAME CONCLUSION: both positive: CF = CF₁ + CF₂(1 − CF₁) both negative: CF = CF₁ + CF₂(1 + CF₁) opposite signs: CF = (CF₁ + CF₂) / (1 − min(|CF₁|, |CF₂|)) WORKED, two rules both concluding the same diagnosis: CF₁ = 0.6, CF₂ = 0.4 CF = 0.6 + 0.4(1 − 0.6) = 0.6 + 0.16 = 0.76 Two moderately confident pieces of evidence combine to 0.76 — MORE than either alone but LESS than their sum, and the result can never reach 1 from finitely many imperfect pieces of evidence. That asymptotic behaviour is the point of the formula. ANOTHER: CF₁ = 0.8, CF₂ = 0.5 CF = 0.8 + 0.5(0.2) = 0.90 CONFLICTING EVIDENCE: CF₁ = 0.7, CF₂ = −0.3 CF = (0.7 − 0.3) / (1 − 0.3) = 0.4/0.7 = 0.5714 THE HONEST CRITICISM OF CERTAINTY FACTORS, which you should be able to state: THEY ARE NOT PROBABILITIES AND HAVE NO AXIOMATIC JUSTIFICATION. The combination formula assumes the pieces of evidence are independent, does not distinguish "no evidence" from "conflicting evidence that cancels", and cannot represent explaining away — the intercausal reasoning from the belief-networks topic. Probability theory does all of this from three axioms, which is why the field moved to Bayesian networks. CERTAINTY FACTORS WORKED WELL IN MYCIN BECAUSE THE RULE SET WAS CAREFULLY HAND-TUNED, not because the arithmetic was sound. ADVANTAGES AND LIMITATIONS — the examinable lists: ADVANTAGES · expertise is PRESERVED — it does not retire or resign · CONSISTENT: the same case gets the same answer every time, where a tired human may not · AVAILABLE everywhere at once, and cheaply replicable · EXPLAINS its reasoning, so it can be checked and taught from · handles UNCERTAINTY explicitly · can COMBINE the knowledge of several experts LIMITATIONS · NARROW: no common sense, and no awareness of the boundary of its own competence. It will answer confidently outside its domain. · BRITTLE: performance degrades sharply rather than gracefully at the edges of the rule set · CANNOT LEARN from experience — every improvement is a hand-edit · the KNOWLEDGE ACQUISITION BOTTLENECK makes them expensive to build and, decisively, expensive to KEEP CURRENT · MAINTENANCE degrades: as a rule base passes a few hundred rules, interactions between rules become impossible to predict, and adding a rule can break behaviour nobody thought was related · no creativity, and no ability to reason about novel situations THOSE LAST THREE ARE WHAT ENDED THE COMMERCIAL BOOM, exactly as the AI history topic recorded. The architecture was sound; what failed was the economics of populating and maintaining the knowledge base by hand.