Rule-based vs case-based vs model-based reasoning

Three ways to reason in an expert system

The exam asks you to give three advantages of each of rule-based, case-based, and model-based reasoning, and to say when each fits. Hold the three paradigms apart by what each *stores* and *reasons from*.

### Rule-based reasoning (RBR) Stores expert knowledge as if-then heuristics mapping symptoms to conclusions. Advantages: 1. Captures expert heuristics directly — the rules *are* the expert's compiled experience. 2. Easily modified — add or delete one rule with no global side effects (the production-system property). 3. Focused, traceable explanation — the rule chain gives a clean why/how account.

Weak when a problem falls *outside* the rules: no first-principles fallback, no causal explanation, brittle at the domain edge.

### Case-based reasoning (CBR) Stores a library of past solved problems (cases) and solves a new one by *retrieve → adapt → apply → save*. Advantages: 1. Encodes historical knowledge cheaply — harvest cases from records instead of interviewing experts (avoids the knowledge-acquisition bottleneck). 2. Avoids repeated search — reuse a past solution instead of re-deriving it. 3. Learns from experience — saving each new case improves the system over time.

Weak at deep explanation, at managing a huge case base, and at defining a good similarity measure.

### Model-based reasoning (MBR) Stores an explicit structural/functional model of how a device *should* work, and diagnoses by comparing predicted vs observed behaviour. Advantages: 1. Robust to novel faults — reasons from the model, so it handles failures no rule anticipated. 2. Causal explanation — explains *why* ("component C's output disagrees with the model"), not just *what*. 3. Transfers across tasks — the same model supports diagnosis, monitoring, and reconfiguration.

Weak when no good model exists, when heuristic expertise matters more than first principles, and because it is computationally expensive. (MBR also inherits a closed-world limit: a fault not in the design model — like a *bridging fault* — cannot even be hypothesised.)

When to use each

Match the paradigm to the dominant knowledge source: rich expert heuristics → RBR; a strong theoretical/causal model (electronics, circuits) → MBR; a deep history of solved cases (law, medicine, help desks) → CBR. Because each one's weakness is another's strength, real systems often go hybrid (e.g. CBR as a fast first check, MBR as the first-principles fallback).