Computational Logic and Human Thinking: How to Be Artificially Intelligent

by Robert Kowalski

The practical benefits of computational logic need not be limited to mathematics and computing. As this book shows, ordinary people in their everyday lives can profit from the recent advances that have been developed for artificial intelligence. The book draws upon related developments in various fields from philosophy to psychology and law. It pays special attention to the integration of logic with decision theory, and the use of logic to improve the clarity and coherence of communication in natural languages such as English. This book is essential reading for teachers and researchers who may be out of touch with the latest developments in computational logic. It will also be useful in any undergraduate course that teaches practical thinking, problem solving or communication skills. Its informal presentation makes the book accessible to readers from any background, but optional, more formal, chapters are also included for those who are more technically oriented.

Genres: Science

334 pages, Hardcover

First published June 27, 2011

Reviews

[Gavin · Rating 3 out of 5]

Nice mission: to teach computer logic to humans to help them think. (Returning logic to its normative roots.) But Kowalski immediately strays from this to also try to build "a comprehensive, logic-based theory of human intelligence". By aiming at both pragmatic self-help and grand, metaphysical, qualitative psychology, it's too ambitious - or rather, appropriately ambitious but using the wrong tools.

(The right tools are unknown but probably include decision theory, statistics, distributed representations, the Bayesian or predictive brain - none of which Kowalski foregrounds. He talks about inferring causes - without using Causal Inference; about doing abduction - without probabilities; about production systems - without the more mature Predictive Processing calculus.)

Kowalski praises a few bad theories, like Fodor's version of language of thought, and Gardner's multiple intelligences. (And Cyc isn't marked as a failure.) But also good theories: dual-process psychology, Sperber's relevance theory.

The best bit is where he links cognitive biases to naive logical rules

The computational interpretation [of dual process theory] is that, when an agent is deliberative, its behaviour is controlled by a high-level program, which manipulates symbols that have meaningful interpretations in the environment. But when the agent is intuitive, its behaviour is generated by a low-level program or physical device...

The logical interpretation of dual process theories is that, when an agent is deliberative, its behaviour is generated by reasoning with high-level goals and beliefs. When the agent is intuitive, its behaviour is determined by low-level input–output associations, even if these associations can also be represented in logical form.

It's also a friendly introduction to more recent logics. Perhaps too friendly - if you think that formal symbols always make things harder to think about, I recommend comparing learning logic from this vs a good semiformal text like Tomassi. The bloat of English compared to symbols is about 20x, and the overheads are impossible to miss.

It is at least what I hoped it would be: a very clear introduction to good old "GOFAI" in all its rigour, grandiosity and narrowness. (There are maybe 600 definitions in this.) I wanted a logician's (or logic programmer's) view on AI, and I got it (from the technical appendices). CL is impressive and authoritative on a small number of tasks, but it's just not generally promising, and hasn't been for a long time. This 2011 book read like a time capsule from the 1970s, before Prolog and Cyc had soured, before the Winter. (I should clarify that inductive logic programming is a live research programme - I'm going to work on it myself - but only in combination with the ruling statistical methods.) I actually don't understand how he can think that this approach is the answer - is it unkind to put it down to decades of sunk cost?

I also thought it might be a more rigorous version of Algorithms to Live By, and I suppose it is, but at the cost of its practicality.