Loading ...
Sorry, an error occurred while loading the content.

40Re: [genetic_programming] Beowulf Clusters

Expand Messages
  • Mattias Fagerlund
    Oct 1, 2001
    • 0 Attachment
      > Our plan is to have one computer hold the population, and farm out
      > evaluations to all computers. So, the faster computers will simply get
      > a higher percentage of the population to evaluate. This has a lot of
      > nice properties, including that we can repeat runs by simply using the
      > same seed, assuming evaluation is deterministic.

      Well, this has, as I understand it, traditionally been avoided because

      1. it scales worse than multiple separate demes, only exchanging about 5%
      of the population (I guess that's 5% each way=10%) cuts down on
      communication

      2. separate demes have been suggested to have beneficial properties over
      one huge population.

      I not really in any position to comment on either claim, but it's
      something to think about.

      > I think the "only exchange info if within 10 generations" thing is
      > particularly dangerous; in one run, populations may be completely
      > separate from almost the start, whereas in another, they'll be sharing
      > info most of the way through. That's a very strange dynamic.

      Yes, this is true, especially since one run might pass the boundry over
      to the next run, and never ever meet the rest of the demes again (they'd
      be on earlier runs all the time). Restarting all runs at the same time
      would take a lot of weird synchronization.

      m
    • Show all 19 messages in this topic