AI is not tossing out all the rules and methods of software development learned over the last 50 years, just many of them.
Rule-based AI systems borrow from rule-based expert system development, which tapped the knowledge of human experts to solve complex problems by reasoning through bodies of knowledge. Expert systems emerged in the 1970s and 1980s.
The knowledge would be represented through if-then-else rules rather than procedural code. Expert systems were considered successful forms of early AI.
Today rule-based AI models include a set of rules and a set of facts, described in a recent account in BecomingHuman/Medium. “You can develop a basic AI model with the help of these two components,” the article states.
Using a machine learning approach, the system defines its own set of rules based on patterns it sees in data. The machine learning system constantly evolves and adapts based on training data streams, relying on models that use statistics. Machine learning models typically require more data than rule-based models.
The author suggests the best projects for rule-based models are when the output is needed quickly or machine-learning is seen as too error-prone. The best projects for machine learning models are those with a fast pace of change and difficult to boil down to a list of set rules.
Machine Learning Programs “Figure Out for Themselves”
Jeff Grisenthwaite, VP of Product, Catalytic
A somewhat similar view was expressed by Jeff Grisenthwaite, VP of Product at Catalytic, a company offering a workflow automation “no code” platform, in an interview published in the Catalytic Blog. “With machine learning, the computer programs can figure out for themselves how to best achieve those goals and can self-sufficiently improve as they intake more data and experience the results of differing scenarios,” he stated.
“With rules-based systems, people define the logic for how the programs make decisions,” he added, using the example of a job recruiting program that disqualifies candidates with less than five years of experience. If a machine learning approach was used to evaluate job candidates, the program would review a large set of training data that includes examples of when candidates were qualified or disqualified. “The program would identify patterns and apply its judgment to new data that comes in, determining a priority ranking of the incoming job candidates,” Grisenthwaite stated.
As to when to use a rule-based approach or a machine-learning approach, Grisenthwaite suggested machine learning is only applicable when thousands of relevant data records are available for making accurate predictions. This could include sales lead qualifications, customer support auto-responses, and situations that have many factors that translate to more columns in a data set.
Machine learning “is better equipped to identify patterns in the data than asking people to both find the patterns and manually develop rules for each of them,” Grisenthwaite stated. An example of this would be algorithms that predict real estate prices, based on a review of historical sales prices and factors including location, square footage and amenities. Also, for rapidly-changing environments such as e-commerce recommendations and sales forecasting, “Machine learning beats out rules-based systems,” he stated.
Rules-based systems are best suited to applications that need lower volumes of data and very straightforward rules. Examples include expense report approvals that define dollar thresholds that require management approvals at various levels, or email routing that uses a list of keywords to determine the destination.
Some systems combine rules-based with machine learning. One Catalytic customer in the advertising business uses a rules-based system to search through a library of answers to prior questions on requests for proposal forms. The responses considered more relevant in that filtered library are then scanned by a machine learning algorithm to predict the best answer to each question.
“Combining rules-based systems with machine learning enables each approach to make up for the shortcomings of the other,” states Grisenthwaite.
“Entire Universe of AI” Can Be Divided Into Rule-Based or Learning-Based
One view is that the “entire universe of AI can be split into these two groups” of rule-based techniques and machine-learning techniques, suggests an account from Tricentis, supplier of a software testing system based on AI.
The authors added, “A computer system that achieves AI through a machine learning technique is called a learning system.” And the goal of a rule-based system is to capture the knowledge of a human expert in a specialized domain and embody it within a computer system.
“That’s it. So let’s regard rule-based systems as the simplest form of AI,” the authors stated, limited by the size of its underlying knowledge base, thus implementing a “narrow AI.”
A dilemma of rule-based systems is the difficulty of adding rules to a large knowledge base without introducing contradicting rules. “The maintenance of these systems then often becomes too time-consuming and expensive,” the authors state. As a result, rule-based systems are less useful for solving problems in complex domains or across multiple simple domains.
Another problem with machine learning systems is that the internal workings of the system cannot be extracted, resulting in a black box, a lack of insight into how the system made its decision. “This is a major problem for many applications,” the authors state. The Equal Credit Opportunity Act, for example, requires that applications for credit must be supplied specific reasons for actions taken.
Dr. Joel Dudley, Chief Scientific Officer, Tempus
A variation of the problems posed by black-box decision-making is the experience of researchers at Mount Sinai Hospital in New York, in applying a learning system to the hospitals’ database of records on some 700,000 individuals. The resulting learning system, called Deep Patient, turned out to be very good at predicting disease. It even appeared to anticipate the onset of psychiatric disorders like schizophrenia, which is difficult for physicians to predict, quite well. “Deep Patient offers no clue as to how it does this,” say the authors, referencing Joel Dudley, former leader of the Mount Sinai team, now chief scientific officer at Tempus Labs, which advances precision medicine through the practical application of AI in healthcare.
“We can build these models, but we don’t know how they work,” Dudley was quoted as saying.