What is the difference between Cased-based Reasoning and Rule-based reasoning?

Question

As stated here, Rule-based Reasoning systems are considered to be "old style" AI that uses rules prepared by humans - as opposed to Neural Networks where machine recognizes pattern i.e. acquires new knowledge and takes decision based upon that.

What I understand about Case-based Reasoning (CBR), it looks at the new cases in light of similar past cases, finds suitable reference cases, evaluates their application on the new case and revises it accordingly, applies it on the new case, and finally stores the case and solution as newly acquired knowledge.

Considering this, can I state that Case-based Reasoning should be considered as "new style" AI? Or is there a gap in my understanding?

---

I wanted to add: If anyone can offer a rough spectrum, that puts different AI approaches on different points depending on how independently they can take decision or how much prone they are to learn from experience (i.e. train their mind) instead of being 'pre-configured', that will also be welcome.

Answer 1

As said in the comments, there may be a part of subjectivity in the answers to this question.

Yet, I think it is fair to say that case-based reasoning mostly belongs to what is often called transductive methods, while rule-based (as well as neural networks and most statistical models) models mostly belong more to what is often called inductive approaches.

In short:

• Transductive methods will not try to induce a generic model from your past observations/experiences, but will simply try to predict/find the answer of a new problem by looking at past experiences. This means you have to maintain a memory of past experience, and then select the most adequate one to answer your new problem. The most well-known representative of such approaches is probably K-NN.

• Inductive approaches will try to induce a "compact" model from your past observations/experiences, and to generalise (make an induction) from those. Once this is done, you can forget your past experiences and only use your learned model to make prediction.

The parallel you can make is between looking up a table of stored input/output values (e.g., to get a p-value), which is transductive, and using a mathematical formula to compute a new prediction, which is inductive.

Lastly, it should be noted that the way you build your model when performing induction largely depends on your applications, available knowledge and needs (there are certainly ways to learn rule-based systems from data).