Expert Systems and AI Approaches to Complex Challenges

UPI Study › Courses › Introduction to Artificial Intelligence › Free lessons › Lesson 2.3 Free with a free account

Expert Systems and AI Approaches to Complex Challenges

Introduction to Artificial Intelligence · Chapter 2: Intelligence, Cognition, and Agent DesignSection 2.3 · lesson 11 of 6914 min · free with a free account

← Evaluating How We Think: Cognitive Function AnalysisNext lesson: Building Blocks of Intelligent Agents →

Earn 3 college credits for Introduction to Artificial IntelligenceRecommended for credit by NCCRS · Computer Science 373 · self-paced · one payment, lifetime access
Enroll in Introduction to Artificial Intelligence · $250 →Download the syllabus (PDF)

In this lesson

How expert systems work, why they hit limits, and alternative approaches.

  • Describe the four main parts of an expert system and how they work together
  • Explain what DENDRAL and MYCIN contributed to the field
  • Explain brittleness and the knowledge acquisition bottleneck
  • Choose between rules, search, learning, and hybrid approaches for a hard problem

Free study notes

Get this lesson's notes by email

The key points and 4 practice questions with answers, in your inbox. One email, free.

One email. Unsubscribe in one click.

Transcript

Everything said in the video, section by section (1,964 words). Read along, or search it.

Opening Question

Capturing Scarce Expert Know-How

A skilled mechanic listens to an engine for a few seconds, and names the problem. A senior doctor glances at a few test results, and knows what to ask next. Expertise like that is valuable, scarce and slow to train. And it walks out the door when the expert retires. In the 1960s and 1970s, AI researchers asked a bold question. Could an expert's know-how be written down, and put into a program?

Definition

Expert-Level Advice in One Narrow Field

The answer was the expert system: a program that captures the specialized knowledge of human experts in one narrow field, and uses it to advise or solve problems at an expert's level. Expert systems are a classic form of symbolic AI, the rule-based approach from section 1.2. They dominated applied AI through much of the 1980s. So let's open one up, meet two famous examples, see why they hit a ceiling, and compare them with other ways to tackle hard problems.

Architecture

Four Parts Around a Car-Trouble Advisor

Whatever its field, an expert system has the same four main parts. A car-trouble advisor makes a good running example. The user asks for advice through the user interface. Behind it, the inference engine does the reasoning, using the knowledge base. The facts reported so far sit in a short-term store beside the engine. Experts and engineers build the knowledge base. And the explanation facility answers questions about the reasoning.

Knowledge Base

Stored Facts and If-Then Rules

Start with the knowledge base. It holds the expertise itself: facts about the field, plus the rules of thumb experts use. Most systems write those rules as if-then statements. Here are three made-up rules. R1: if the engine doesn't crank, and the headlights are dim, then the battery is probably weak. R2: if the engine cranks but won't start, and the fuel gauge reads empty, then the tank is probably empty. R3: if the battery is probably weak, and it's more than 4 years old, recommend replacing it. These rules came from mechanics, not from data. And the knowledge base only stores them. It doesn't do any thinking. Writing knowledge down formally comes back in section 6.7.

Inference Engine

Chaining Rules to Reach Advice

The inference engine does the reasoning. It takes the rules and the facts of the current case, works out what follows, and decides what to ask next. Say the user reports that the engine won't crank, and the headlights are dim. The engine sees that R1 applies, and concludes the battery is probably weak. That new conclusion lets R3 apply, as soon as the user adds that the battery is five years old. The advice: replace it. Rules link like steps in a chain, one rule's conclusion feeding the next. If the knowledge base is the mechanic's experience, the engine is the mechanic's thinking.

User Interface

A Question-and-Answer Conversation

The user interface is how people talk to the system. Most expert systems work as a conversation. The system asks questions. The user answers yes, no, a number, or unknown. Then the system reports its conclusion and advice. A good interface uses the vocabulary of the field, and asks only relevant questions, because the user is usually a professional or a customer, not a programmer.

Explanation Facility

Answering Why and How

The explanation facility lets the system justify itself. Two questions matter most. Why are you asking me this? And how did you reach that conclusion? Mid-conversation, the user asks why the system wants to know about the headlights. It answers: rule R1 needs both facts. After the conclusion, the user asks how. It answers: R1 applied, because you said the engine doesn't crank and the headlights are dim. Why it matters: users trust advice they can check. Experts can spot and repair faulty rules. And in medicine, finance and law, a person stays responsible for the decision, and must defend it.

Quick Check

Updating a Tax-Advice System

Quick check. A tax-advice expert system needs updating after a new tax law. Working with a tax expert, the team writes the new law into the system, as if-then statements. Its advice changes as a result. Which part of the expert system did they change? Pause, and decide. The knowledge base. The rules are the stored expertise, so new tax rules go there. The inference engine is tempting, because the system's advice changes. But the engine is the general reasoning program, and it stays the same. It simply reasons with the new rules. Keeping knowledge and reasoning apart is exactly what makes this kind of update possible.

DENDRAL

Inferring Molecular Structure From Spectra

Two systems built at Stanford did more than any others to establish the field. DENDRAL began in 1965, and is often called the first expert system. Its task was to work out the structure of an unknown organic compound, from mass spectrometer data. That instrument breaks molecules into fragments, and records the pattern. The number of possible structures explodes as molecules grow, so trying them all is hopeless. So expert chemists' rules marked which structures were plausible. A generator built candidates only among those. Each candidate's predicted pattern was compared with the real readings, giving the chemist a short ranked list. The lasting lesson: in a narrow domain, specific knowledge matters more than clever general reasoning tricks.

MYCIN

Certainty Factors and Antibiotic Advice

MYCIN, from the early 1970s, diagnosed serious bloodstream infections and meningitis, and recommended antibiotics. At its mature stage it held about 600 rules. Three features made it a model. First, it handled uncertainty. Each rule carried a certainty factor, a number for how strongly the evidence supported its conclusion. Second, it explained itself, answering why and how. Third, its designers kept the knowledge apart from the reasoning program. Strip out the medical rules, and an empty framework was left, called EMYCIN, that could be filled with rules from other fields. In evaluation studies, MYCIN's advice was rated comparable to human specialists. Yet it was never used routinely in patient care. Reasons often cited: open questions about who is responsible when a program gives medical advice, the limited computers of the day, and long question sessions doctors would have had to sit through. Still, the two systems set off a boom. By the 1980s, companies used expert systems to configure computer orders, diagnose equipment faults, and check credit applications.

Strengths

Consistent, Available, and Explainable

Where do expert systems shine? They're consistent: the same facts always give the same advice, with no fatigue and no bad days. They're always available, serving many users at any hour, at a low cost per use. They preserve scarce expertise after the expert leaves. They're explainable: every conclusion traces to rules people can read. And they need little data, working from expert knowledge alone.

Brittleness

Solid Until It Shatters

Inside its narrow field, a good expert system can match human experts. At the edges, it fails sharply. That's called brittleness. Like glass, it's solid until it's pushed past its limit, and then it shatters instead of bending. The causes are built in: no common sense, no awareness of where its knowledge ends, and no way to learn from its mistakes. Our car advisor was written for gasoline engines. Describe an electric car that won't start, and it has no relevant rules. Worse, it doesn't know that. It may keep asking about the fuel gauge, and still give confident but meaningless advice. A person would say: that's outside my specialty.

Knowledge Acquisition

Experts and Knowledge Engineers in a Loop

Building the knowledge base is the hardest part. Two kinds of people do it: the domain expert, who has the knowledge, and the knowledge engineer, an AI specialist. The engineer interviews the expert, watches them work, and studies their cases. Then they write rules, test the system on cases, and compare its advice with the expert's. Wherever they disagree, the rules get fixed, and the loop runs again, many times over. It's so slow and costly that it became known as the knowledge acquisition bottleneck.

Why It Is Slow

Hidden Knowledge and Interacting Rules

Why so slow? Hidden knowledge: experts often can't explain how they know. A mechanic just knows a certain sound means a worn bearing. Disagreement: two doctors may prefer different treatments for the same case. Time and cost: months of sessions with busy, highly paid experts, for a few hundred rules. Interacting rules: a new rule for older cars can quietly override R3. And keeping up: a new drug or a new tax law forces revisions by hand. An expert system can't learn on its own, so every improvement goes through this human pipeline. With brittleness, that's why the boom slowed in the late 1980s.

Four Approaches

Rules, Search, Learning, and Hybrids

Expert systems are one answer to the question: how should a machine tackle a hard problem? Three other families sit beside them. Rules write experts' knowledge as if-then statements. They work best when knowledge is explicit and stable, explanations are required, and data is scarce. But they're brittle and costly. Search explores possible options on the way to a goal. It works when the goal is clear and options can be scored. But the options can explode. Learning finds patterns in many examples, when rules are too hard to state. But it needs data, and is hard to explain. And hybrids combine tools, at the cost of more complex design.

Choosing a Tool

Questions to Ask About the Problem

A practical way to choose is to ask about the problem before choosing the tool. First: can experts state the knowledge as clear rules? If yes, build rules: an expert system. If not: is there a clear goal, and a huge number of options to explore? If yes, use search. If not: are there many past examples, but no clear rules? If yes, use learning. If every answer is no, gather expert knowledge or data first. And many real problems answer yes more than once. Then combine the tools in a hybrid.

Worked Examples

Loans, Routes, Digits, and DENDRAL

Try the questions on four hard problems. Checking a loan application against written lending rules: the knowledge is explicit, and each decision must be explained. Rules. Finding the fastest delivery route through 20 stops: a clear goal, and every route can be scored. Search. Recognizing handwritten digits: nobody can write the rules, but thousands of labeled examples exist. Learning. And DENDRAL itself was a hybrid. Expert rules pruned the options, and a generator searched what remained.

Quick Check

A Hospital Pharmacy Needs Two Tools

Quick check. A hospital pharmacy wants software that blocks any order breaking its written dosing rules, and explains why. It also wants to flag unusual prescribing, using two million past orders, where nobody can state the rules for unusual. Which approach fits best? Pause, and decide. Both kinds, combined. The dosing check has explicit rules that must be explained, so rules fit that part. Spotting unusual prescribing has lots of examples but no clear rules, so learning fits that part. Written rules alone is tempting, since the dosing rules matter most. But rules can't capture what nobody can state. The problem says yes to two questions: a hybrid.

Common Mistakes

Four Misreadings of Expert Systems

Four mistakes to avoid. First, thinking an expert system is smarter than any person. It's narrow, and it can't tell when it's out of its depth. Second, equating brittle with inaccurate. A brittle system can be extremely accurate inside its field. The problem is how suddenly it fails outside. Third, treating rule-based systems as obsolete. Rules remain right where decisions must follow written regulations, and they live on inside many hybrids. And fourth, picking an approach because it's fashionable. Start from the problem: how clear the rules are, how many options exist, and how much data there is.

Key Takeaways

Parts, Pioneers, Limits, and Choices

To sum up. An expert system has four main parts: a knowledge base, an inference engine, a user interface, and an explanation facility. DENDRAL showed that domain knowledge plus systematic generation solves hard problems. MYCIN showed how to handle uncertainty, and explain advice. Brittleness and the knowledge acquisition bottleneck limited expert systems. Rules, search and learning each suit different problems, and real systems are often hybrids.

Reflection

Expertise That Is Hard to Put Into Words

Here's something to think about. Pick a job where expertise matters, and is hard to put into words. Which parts of it could be written as rules? Which parts would need learning from examples? And how would the system know when a case is outside its field?

About Introduction to Artificial Intelligence at UPI Study

Introduction to Artificial Intelligence (Computer Science 373) is a self-paced online course of 69 video lessons in 9 chapters, from intelligent agents and search to game strategy, neural networks, logic and reasoning, machine learning, language, robotics and AI ethics. It carries a NCCRS credit recommendation for 3 semester credits.

Every lesson transcript is free on pages like this one, and the videos are free with a free UPI Study account. Enrolling adds a practice quiz after each lesson, the graded work and final exam that earn the credit, lifetime access, and an official transcript on completion. UPI Study has helped over 130,000 students with college credits. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.

Enroll now → Course details Download the syllabus All Computer Science courses Every course with PRO →

Compare: AI vs Current Trends in CS and IT

Why take it here

Why students take Introduction to Artificial Intelligence with UPI Study

How machines search, reason and learnIntroduction to Artificial Intelligence covers how AI systems are built: intelligent agents, search and heuristics, game strategy, logic, neural networks, machine learning, language and perception, and the ethics of AI. It is a common upper-level course in computer science degrees.
Credit at a fraction of the priceThe same 3 credits can cost $750 or more at a US college. Here it is one payment, self-paced, with lifetime access.
Finish your degree soonerTransfer credits reduce the courses you still owe, so you can graduate earlier or lighten a heavy semester.
Strengthen an applicationA completed college-level course on an official transcript shows admissions and scholarship committees you can do the work.
Study on your own scheduleNo start dates, no deadlines. Watch the lessons, practice, and take the final when you are ready.
130,000+students helped with college creditsACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.

Where the credit goes

Universities that work with NCCRS

Introduction to Artificial Intelligence carries a NCCRS credit recommendation. These universities work with NCCRS; some are UPI Study's signed partners. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.

Thomas Edison State UniversityThomas Edison State UniversitySigned partnerUniversity of the PeopleUniversity of the PeopleSigned partnerCharter Oak State CollegeCharter Oak State CollegeSigned partnerStrayer UniversityStrayer UniversitySigned partnerColumbia Southern UniversityColumbia Southern UniversitySigned partnerIndiana TechIndiana TechSigned partnerPenn StatePenn StateWorks with NCCRSSouthern New Hampshire UniversitySouthern New Hampshire UniversityWorks with NCCRSSUNYSUNYWorks with NCCRS +many more cooperating universitiesFind my college →

Find my college → Not listed? Ask Athena in the corner of this page, or book a 15-minute call.

FAQ

Questions students ask us

About Expert Systems and AI Approaches to Complex Challenges

What are the main components of an expert system?
An expert system has four main parts: a knowledge base, an inference engine, a user interface, and an explanation facility. The knowledge base stores experts' facts and if-then rules, the inference engine reasons with those rules and the facts of the current case, the user interface runs the question-and-answer conversation, and the explanation facility tells users why a question is asked and how a conclusion was reached.
What were DENDRAL and MYCIN?
DENDRAL and MYCIN were two pioneering expert systems built at Stanford. DENDRAL, begun in 1965 and often called the first expert system, worked out the structure of unknown organic compounds from mass spectrometer data. MYCIN, from the early 1970s, diagnosed serious bloodstream infections and meningitis and recommended antibiotics using about 600 rules with certainty factors. Its advice was rated comparable to specialists, yet it was never used routinely in patient care.
What is the knowledge acquisition bottleneck?
The knowledge acquisition bottleneck is the slow, costly work of getting expertise out of human experts and into an expert system's rules. A knowledge engineer interviews and observes a domain expert, writes rules, tests them on cases, and repeats many times. It is hard because experts often cannot explain how they know things, experts disagree, new rules can interact with old ones, and every update must be made by hand.
Why are expert systems called brittle?
Expert systems are called brittle because they can perform at expert level inside their narrow field yet fail sharply at its edges. They lack common sense, cannot tell where their knowledge ends, and cannot learn from mistakes. A car advisor written for gasoline engines, asked about an electric car, may keep asking about the fuel gauge and give confident but meaningless advice. Brittle does not mean inaccurate inside the field.

About the Introduction to Artificial Intelligence course

How many college credits is Introduction to Artificial Intelligence?
Introduction to Artificial Intelligence (Computer Science 373) carries a NCCRS credit recommendation for 3 semester credits. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.
How much does Introduction to Artificial Intelligence cost and what is included?
$250, paid once, with lifetime access: all 69 video lessons, a practice quiz after each, the graded work and final exam, and an official transcript when you finish. Students who want several courses choose UPI Study PRO instead.
Which universities accept UPI Study credits?
UPI Study has signed partner agreements with universities such as Thomas Edison State University, University of the People and Charter Oak State College, and works with universities such as Penn State, SNHU and SUNY. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities. Use "Find my college" on this page to check yours.
How long does Introduction to Artificial Intelligence take?
It is self-paced with no deadlines. The 69 lessons total about 16 hours of video; most students add reading and practice time and finish in a few weeks to a few months, on their own schedule.
How are the exams done?
Quizzes and assignments are online in the course. The final exam is taken online with proctoring; a person on the UPI Study team checks assignments, usually within two days.
How do my credits reach my college?
When the course shows 100% complete, request the official transcript from your dashboard. It is sent to the college you name, and you can send it to more than one. Check your college first with "Find my college" on this page, or ask Athena.

Ready for the full course? Earn 3 credits in Introduction to Artificial Intelligence.

69 video lessons, a quiz after each one, and the graded work that earns the credit. Recommended for credit by NCCRS.

Enroll in Introduction to Artificial Intelligence →Download the syllabusEvery course with PRO →
Introduction to Artificial Intelligence
3 credits · NCCRSEnroll →