Rule-Based Reasoning

Rule-based reasoning is a symbolic artificial-intelligence technique that represents knowledge as “if… then…” rules and uses an inference engine to derive conclusions from a set of facts. Unlike a neural network, the knowledge is explicit and readable. In its classic form an expert writes it; there is also a family of techniques, rule induction, that learns the rules themselves from labelled data —using algorithms such as C4.5 or RIPPER— even though the result remains a readable set of rules. It is the paradigm behind the classic expert systems.

How it reasons: rule chaining

The inference engine chains the rules in one of two directions. In forward chaining, driven by the data, it starts from the known facts, fires the rules whose conditions are met and adds their conclusions as new facts, until it reaches a goal. In backward chaining, driven by goals, it starts from a hypothesis and works back through the rules, breaking it into subgoals until it finds facts that support it; this was the mechanism of the medical expert system MYCIN, which also chained its rules with certainty factors — a numerical confidence measure between -1 and 1, not strictly binary logic.

What it is not

Two common confusions are worth undoing. Backward chaining is not “backpropagation”: that is the algorithm that trains neural networks by adjusting numerical weights with the gradient of the error, not an inference over rules. And it is not “backtracking”: that is a search strategy that undoes assignments when it hits a constraint, typical of constraint-satisfaction problems. Both terms belong to other fields.

Components, virtues and limits

A rule-based system consists of a knowledge base (the rules), a working memory (the facts of the case) and the inference engine that combines them. Its great virtue is transparency: every conclusion is traceable to the rules that produced it. And a mistaken idea is worth correcting: these systems are not “self-evolving”: they do not modify their own rules on the fly, and in their most common form an expert maintains them. Not all of them are deterministic — MYCIN, as seen above, reasoned with certainty factors — and not all their rules are born without data, as rule induction shows. Their limits are rigidity in the face of unforeseen cases and the cost of maintaining large rule bases. The current line is neuro-symbolic systems, which combine rules with neural networks; but there the ability to learn comes from the neural component, not from the rule paradigm.