View Source Document

Chris Pressey.md

Commentary by Chris Pressey

This work is distributed under a CC-BY-ND-4.0 license, with the following explicit exception: the ratings may be freely used for any purpose with no limitations.

Artificial Intelligence

Artificial Intelligence, 3rd ed.

Mostly GOFAI ("Good Old-Fashioned Artificial Intelligence") of course - goal search strategy type stuff, truth maintenance, and so forth.

It is full of snappy maxims like

"Deduction Systems May Run Either Forward or Backward"

Chapter 14 is about backtracking an truth maintenance, but I don't follow the example well. I assume Prolog implementations have absorbed as much truth maintenance optimization as is possible in the setting of Prolog. For surely in the general case there are limits to how much truth your goal-searching algorithm can productively maintain.

Chapter 15 is about planning, which includes the situation calculus and frame axioms, and describes Green's trick, which is a way of capturing a trace while a plan is searched for:

Tracing the situation history is a tedious, error-provoking chore for problems involving many operations. Consequently, the usual practice is to use something called Green’s trick. The basic idea is to add an extra term to the negation of the desired result. This extra term, called the answer term, exploits the ordinary unification apparatus to keep track of situation information. Instead of

ON(B, Table, s)

you write

ON(B, Table, s) V Answer(s)

Prolog programming for artificial intelligence, 3rd ed.

Mostly GOFAI ("Good Old-Fashioned Artificial Intelligence") of course - goal search strategy type stuff.

15.4 shows how to generate a proof tree as you go in Prolog. (Which is related to, but not the same as, Green's trick. For that, see [Artficial Intelligence, 3rd ed.][] by Winston.)

23.5 talks about "pattern-directed programming" which basically means a loop that picks all applicable rules, resolves conflicts in that selected set, applies them, repeats. This includes Prolog and term rewriting systems at least. Then a theorem prover is shown using this approach. This makes sense because rules of inference can be picked just as easily as rewrite rules can.

Building Problem Solvers

More about truth maintenance.

Sussman anomaly - Wikipedia

.

soft question - A clear map of mathematical approaches to Artificial Intelligence - MathOverflow

.

CS229br Foundations of Deep Learning (aka Topics in the Foundations of Machine Learning)

.

Universal approximation theorem - Wikipedia

.