[Prev][Next][Index][Thread]
RE: AI and mice
A lot of the ANN approach is like using Monte Carlo simulation to evaluate
pi - a lot of hard work to get a result which was obvious. In a case like
maze solving, the problem is deterministic and easily 'formalised'. To
produce exemplars and reward functions for ANN applied to maze solving as a
whole, or to imitate a rat is not conducive to success in competition!
Where learning and training will really pay off is in learning manoeuvres,
rather than strategies. What the mouse builder does is to set the speed
higher and higher until the mouse eventually crashes, then backs off a bit.
A mouse could do that itself during practice - a crash is easily detected.
Dave Woodfield had a neat strategy a while ago where a mouse entered a
corner and executed an 'open loop' manoeuvre in which the steering was
'waggled blind' according to a pre-learned (via David) sequence of steering
angles. Feedback was only used to sort things out after the corner. That's
another one the mouse could learn adaptively.
There are hosts of nonlinear and adaptive stragegies which give untold
advantages. A tiny subset of these has received attention with the
buzzwords 'Neural' and 'Fuzzy' - and the baby has usually gone out with the
bathwater. By chasing a 'brand name' algorithm you are no more likely to
gain an advantage than by wearing Nike shoes at the contest.
Cheers
John
Somebody asked about neural networks and maze solving. I passed the query
onto one of our new staff (Craig Saunders - he was the one running the
timing computer at the competition) and this is what he had to say.
As for using NN's (or other machine learning tasks) for maze solving the
simple answer is yes - it has been looked at before. As far as I know (and
this may well be wrong) there isn't a great deal of information regarding
this out there. There seems to be two approaches to maze-solving with
NNs:
1) Can we construct an Artificial Neural Net (ANN) to model the behaviour
of an animal that solves a maze (subtly different from actually
getting an ANN to solve it itself). There was something called RATNET
(or similar) which looked at this about 5 years ago (I think). Anyway
people in this school of thought tend to be biologists or
psychologists using ANNs to emulate behaviour.
2) The ML approach to solving mazes is usually based around some type of
re-inforcement learning. There are direct re-inforcemnet learning
methods or you can us RF-learning techniques in conjunction with ANNs,
genetic algorithms etc.. Again, I vaguely remember talks from a
couple of years ago where people used a NN based algorithm to deal
with time-series data which solved a maze...
These ideas are vague, but you (hopefully) get the gist. A search engine
should delve something up - if not I can probably dig up a few links. If
your going to use an "intelligent" method for maze-solving, then
reinforcement learning is definately the place to start (in my opinion
anyway).
Of course, another approach would be to get an "intelligent" algorithm to
learn how to interpret the sensors and actually move the mouse. E.g. the
algorithm would "learn" which adjustments are best in order to move the
mouse in a straight line, how best to take corners, etc..
I was going to have a look at this before the next competition (if anyone
would like to donate a mouse! - or alternatively I'll have to learn
something about electronics), you never know, I might have the time....
Craig.
----------------------------------------------------------------------------
---
| Craig Saunders | e-mail :
C.Saunders@dcs.rhbnc.ac.uk |
| Dept. of Computer Science | Phone : (+44) 1784 443437
|
| Royal Holloway and Bedford | Fax : (+44) 1784 439786
|
| New College, University of London | Mobile : 07801 478748
|
|--------------------------------------|------------------------------------
--|