Self reinforcing strategies

  • 1
  • Article
  • Updated 8 years ago
I did wonder whether there might be an issue until 'all builds including bad ones' are synthesised with the ordering for the strategy markets.

Lets take for example the 'clean dotplot' strategy

It's currently got a good rating according to the strategy market for how successful designs that fit that do in the lab.

But 'because' that seems to make sense to us - most of the designs we have made have had a fairly clean dotplot and very few synthesised designs will have had a very different dotplot.

So whilst the few grey dotplot builds may (by chance or by inherent quality of grey dotplots) have been bad, they are probably not a large enough proportion of the sample - and that will tend to reinforce itself as cleaner dotplots will tend to get voted on and therefore synthesised, and so clean dotplots will tend to get good marks (because they have humans designing them so are certainly better than random) which results in the system reinforcing the stats saying that clean dotplot is best.

I'm wondering if that's been considered - and I think that means that until the 'big pile of unsynthesised designs' (which will include lots of grey dotplot that may or may not be rubbish) have been made you probably can't do more than take the ordering values as assessment of what the players tend to build their designs to look like.

But that might be good enough?
Photo of Edward Lane

Edward Lane

  • 139 Posts
  • 8 Reply Likes

Posted 8 years ago

  • 1
Photo of Edward Lane

Edward Lane

  • 139 Posts
  • 8 Reply Likes
ok found an interesting example

this strategy (which penalises what Eli called double self turning GC pairs) http://eterna.cmu.edu/eterna_page.php...

The results seem to show that no design scored anything other than zero ?

So I think that means no design that has ever been voted for and thus synthesised has ever had 'too many' double self turning GC pairs.

is that because they 'look' bad based on the assumption that this strategy is correct?