Reactive planning and nouvelle AI


next up previous contents
Next: The middle way
Up: Introduction - some problems with classical AI
Previous: Introduction - some problems with classical AI
Back: to main list of student notes

Reactive planning and nouvelle AI

PopBeast goes to one extreme - trying to build a complete plan of action before doing anything. Some people believe that it's time to experiment with the opposite approach - thinking ahead as little as possible, and reacting directly to the real world wherever possible. This is one of the differences between ``classical AI'' and what is sometimes called ``nouvelle AI''.

Trying to generalise over the subject is risky. But I think it's fair to say that the typical nouvelle AI approach is this: conventional AI has had such little success that we need something completely different. Most of what's needed in a natural intelligence was placed there by evolution long before the human brain evolved (compare the two billion years life has existed with the history of the human race). Conventional AI has concentrated on rational intelligence, printing the postage stamp on top of the Empire State Building without worrying about the building itself. What we should be doing is to work our way up the evolutionary tree, making a complete creature at each step and then building the next on top of the last. So we should start by trying to mimic insects. The result won't be able to solve chess problems or prove theorems, but it should be able to wander around the lab without getting stuck under chairs, running out of energy because it's failed to notice that its batteries are running low, or falling over and being unable to right itself.

Once we have a working insect, we can work up from there. Eventually, we shall find that in the course of building competance to operate in the real world, we have most of what's needed for rational thought. For example, spatial problem solving may fall out as a byproduct of visual navigation. The last few steps will then be easy.

The table below summarises the differences between an extreme classical AI position and an extreme nouvelle AI one. Most workers will not be so extreme, and will end up somewhere in between. Incidentally, there are many dimensions of variation between different AI techniques. The typical classical approach and typical nouvelle approach differ in several of these. This is not evident from the table below.


next up previous contents
Next: The middle way
Up: Introduction - some problems with classical AI
Previous: Introduction - some problems with classical AI
Back: to main list of student notes



Jocelyn Ireson-Paine
Thu Feb 15 00:10:39 GMT 1996