next up previous
Next: Preamble
Up: Conventional AI: Evans' Analogy Program
Previous: Conventional AI: Evans' Analogy Program

Introduction

I'll set your first essay, and give a tutorial, on Representations and Evans' Analogy Program.

Why? First, a question. How does diamond differ from graphite? Graphite is soft enough for pencil leads; it's opaque, it's a good conductor of electricity, and (as charcoal) it makes an excellent fuel. Diamond is the hardest natural substance known; it's transparent, it doesn't conduct electricity, and not even the most desparate barbecue fanatic would throw his wife's wedding ring into the coals. Yet both are made from the same atoms.

The difference is form. In graphite, the atoms are arranged in layers that slip easily over one another, allowing the free passage of electrons, and coming off in streaks onto paper. In diamond, the atoms form a tetrahedral structure; it has no ``give'', and you must work very hard to detach an atom.

This is obvious to any chemist. But until 1860 or so, the difference was inexplicable. Although Dalton had put the idea of atoms combining in definite ratios, the notion of shape and arrangement of atoms was missing; no-one seems to have considered it until Van't Hoff proposed that carbon atoms have a definite shape, with bonds pointing outwards as from the centre of a tetrahedron.

There is an analogy with psychology. Programmers know that the same piece of information can be represented in many different ways: numbers in base 10, base 2, or as Roman numerals; dates as day/month/year or as days since January 1st 1890; images as arrays of intensity points or as a set of classified objects; geometric facts as drawings on graphs or co-ordinate equations ....

These choices are not arbitrary. The way something is represented on a particular information-processing system crucially affects the types of operation we can perform naturally and quickly. For instance, if we represent dates as days since a base date, it's easy to find the time between two dates, but hard to tell which month any date falls into. With the day/month/year representation, the converse is true.

So representation should be a crucial idea in AI, in computing, and in psychology. But until the advent of computational models it has been largely neglected. What made us realise its importance was the technical difficulty of programming computers. How many psychological models contain boxes labelled with names like Thought, Memory, Inference? What passes from one box to another? Information --- but these models rarely consider its form.

So I'm going to start AI by talking about representations, and about a program, Evan's Analogy Program that is built round the notion of a good representation.


next up previous
Next: Preamble
Up: Conventional AI: Evans' Analogy Program
Previous: Conventional AI: Evans' Analogy Program



Jocelyn Paine
Tue Jun 3 11:33:37 BST 1997