has been around 15-20 years now and AFAIK
it hasn't turned up anything that you could fairly say proves we've got the principle cracked. I'm afraid I'm not impressed by Norns
is obviously capable of great things, but I don't think we've properly translated the mechanism into sillicon
yet. This is mainly because most ALife
researchers have a very simple and crude understanding of it. This runs along the lines of:
- Start with a population of random candidate solutions.
- Assign each one a fitness score according to some pre-defined criteria.
- Increase frequency of high-fitness solution paramters by breeding them or copying with mutation.
- repeat as necessary.
This looks like it should work. The only problem is the `repeat as necessary
' part. The search space
defined by the variable set can be prohibitively large for most non-trivial problems. The standard solution of course is to throw more processing speed at the problem
whereas what would really
be useful would be to go back and have a proper look at how biology
does it. Until this happens AL
is in danger of falling into the same trap of AI
, i.e. producing lots of systems which are very interesting but are almost completely disparate and don't really go anywhere.
BTW, this whinge does NOT include neural networks which do work well.
I've just re-read this after noticing it pick up negative xp
. I still agree with what I said, but to really justify
it properly would require a much bigger node. Maybe even an alife metanode
. Instead I'll try to state succintly
where I'm coming from...
That's it! I just want to see ALife stuff progress as I know it can...