While expert systems have been one of the earliest forms of successful AI, they have limitations in terms of:
- (i) knowledge acquisition – how to get the sufficient knowledge captured with acceptable expert effort;
- (ii) verification – how to ensure rules are accurate;
- (iii) brittleness – difficulty in generalisation beyond fixed domains; and
- (iv) meta-knowledge – ability to reason about their own processes.
Many expert systems now include neural networks for decision making (e.g. spotting abnormalities in X-rays) and may use large-language models as part of their user interface. This can help with (i) and (iii), but makes (ii) and (iv) even more difficult.
Used in Chap. 18: page 277