Learning Systems

[David Crocombe]

In my last job I used to service a tape recorder synchroniser
that would learn the performance of each tape recorder so 
that it could control it in the most efficient way.
It would do a few fast forwards, rewinds, stops and measure
the dynamics. Each time it would get a bit better at
controlling the machine so that if for example it wanted to
go to a particular position on the tape it would know how
far before that point it had to go into braking from fast
forward to be able to stop slightly ahead of the required
point and then drop into play mode to find the exact point.
This was very impressive to watch its performance get
noticably better as it learnt the dynamic performance.

It would be interesting to play with this kind of learning
in relation to cars. The vehicles could then be self tuning
depending on environment, load, parameter drift, wear or
whatever. I guess traction control and antilock brakes are
a step in this direction of making cars more intelligent.

David Crocombe,
Sydney,
Australia.


---------------------------------------------------------
                  Reply Seperator
---------------------------------------------------------


>There appears to exist a number of sensors to build
>an excellent self tuning EFI, yet the only sensor that
>anyone talks about for ignition timing is a knock
>sensor that just tells me - oops too much -
>self destruct sequence initiated.

/cut/

The only thing stopping you having a computer tweak the
engine parameters for you is that you have to have
*accurate* *results* from your measuring system.
It seems to me with an accurate performance measure-
ment system, anything is possible, but building an
*accurate* system is a bit of an ask.

/cut/

Use some fuzzy logic to try to improve things over
multiple acceleration runs, hey presto, self tuning car.

/cut/

record multiple instances of the same speed/RPM/TPS for
different spark/fuel parameters, historical EGO info...

My idea is to have this running for a race track environment,
where the car is doing multiple acceleration and speed runs,
repetitively, and I think given a baseline spark/fuel curve, it
could quite quickly approach the perfection asymptote...