Artificial Intelligence

02Capturing Human Expertise in Simple Rules

We often rely on highly trained specialists to solve our most pressing problems, from diagnosing a strange engine noise in our car to identifying a rare medical condition. What if we could bottle that hard-earned expertise, preserve it forever, and hand it to anyone who needs it? This is exactly the promise of rule-based expert systems, one of the earliest and most commercially successful branches of artificial intelligence. In this chapter, we explore how computer scientists learned to interview human specialists, extract their knowledge, and translate it into a language that machines can execute flawlessly. At the incredibly beating heart of every expert system lies a remarkably simple concept: the IF-THEN rule. While human decision-making feels incredibly complex and nuanced, Negnevitsky reveals that much of professional expertise can actually be distilled into these straightforward conditional statements. Consider a seasoned mechanic examining a car that refuses to start. The mechanic does not consciously run complex calculus equations in their head; instead, they rely on a chain of logical deductions built over years of experience. They might think, "If the engine does not turn over, and the headlights are dim, then the battery is likely dead." This is a classic IF-THEN rule. The "IF" part represents the condition or the evidence we observe, while the "THEN" part represents the conclusion or the action to be taken. Building an expert system requires two main components: the knowledge base and the inference engine. You can think of the knowledge base as a massive library containing hundreds or even thousands of these IF-THEN rules. It is the repository of human expertise. However, a library is entirely useless without a librarian to search through the books and find the right answers. This is where the inference engine comes into play. The inference engine is the "brain" of the expert system. It is a computer program that actively searches the knowledge base, matches the rules against the current data, and logically infers new information to reach a final conclusion. To understand how the inference engine navigates this vast sea of rules, we must look at the two primary methods of reasoning: forward chaining and backward chaining. These two approaches mirror the different ways humans tackle problem-solving in everyday life. Forward chaining is entirely data-driven. It starts with the known facts and works forward to see what conclusions can be drawn. Picture an investigator walking into a chaotic crime scene. They gather fingerprints, secure the weapon, and interview witnesses. They take all this scattered data and push it forward through their internal rule base to eventually identify a suspect. In a computer system, forward chaining looks at the available data, fires any rules whose "IF" conditions are met, generates new facts, and repeats the process until no more rules can be fired. This method is incredibly useful for tasks like system monitoring, where data is constantly flowing in, and the system needs to trigger an alarm if a dangerous combination of events occurs. On the flip side, backward chaining is goal-driven. It starts with a specific hypothesis or a goal and works backward to see if the available evidence supports it. Returning to our crime scene analogy, backward chaining is like a detective who already has a prime suspect in mind. The detective's goal is to prove the suspect's guilt, so they work backward, looking specifically for the evidence—financial motives, lack of an alibi, or physical clues—that would validate their hypothesis. In the medical field, a doctor often uses backward chaining. They might suspect a patient has a specific infection based on a quick glance. The doctor will then ask targeted questions and order specific lab tests to confirm or deny that initial hypothesis. Expert systems designed for diagnosis almost always rely on backward chaining because it is much more efficient to ask the user a few targeted questions rather than demanding they input every single fact about their life before reaching a conclusion. One of the most famous historical examples highlighted in the study of expert systems is MYCIN, a program developed in the 1970s to diagnose blood infections and recommend antibiotic therapies. MYCIN was a revolutionary achievement because it proved that a machine could perform at a level equal to or better than human specialists in a highly complex domain. The system worked by asking the attending physician a series of questions about the patient's symptoms and lab results. Using backward chaining, MYCIN navigated through hundreds of rules regarding bacterial infections. However, MYCIN also introduced a critical innovation: the ability to handle uncertainty. Human experts rarely deal in absolute certainties. A doctor might say, "Given these symptoms, I am 80% sure this is a streptococcus infection." To capture this reality, developers introduced certainty factors into expert systems. A certainty factor is a number that represents the level of belief in a particular fact or rule. It allows the system to accumulate evidence from multiple, sometimes conflicting, sources and calculate a final confidence level for its recommendation. This addition made expert systems significantly more robust and human-like in their reasoning. Despite their incredible success in specialized fields, rule-based expert systems also possess notable limitations. They are notoriously "brittle." If a situation arises that falls even slightly outside their programmed rules, they cannot adapt or use common sense; they simply fail. Furthermore, the process of extracting rules from human experts—a process known as knowledge engineering—is incredibly time-consuming and difficult. Experts often struggle to articulate exactly why they make certain decisions, relying heavily on intuition that defies simple IF-THEN logic. This bottleneck in knowledge acquisition signaled to researchers that while rigid rules are excellent for capturing structured expertise, they are insufficient for replicating the fluid, adaptable nature of true human intelligence. To build smarter systems, machines needed to learn how to operate not just in black and white, but in the infinite shades of gray that define the real world.

Continue reading with LeapAhead app

Full summary is waiting for you in the app

Continue reading