Migration and Slow Islands

[michaelgarvie]

Talking about this cool stuff under the "server not working" header was getting on my nerves!! So here we go.

OK, safemode suggested using increased team migration for slower islands only . In effect this would simulate one larger slow island. This big island wouldn't have a faster generation rate, just a larger population. It is one of those GA "black magic" issues as to what the size of a population should be. What could be done instead is that these slow islands could all have smaller population and the island cluster would then be comparable to other faster islands in population size and generation rate. I see how this seems like it could make it more likely for such an island cluster to come up with a fitness improvement to the current elite. However this moves me on to the second point raised by safemode.
.
Slow islands don't necesarily contribute anything to the evolutionary run. I will certainly disagree here. Think of a straighforward bruteforce search algorithm such as distributed.net. Slow nodes in such a cluster do contribute an amount proportional to their speed. However the way such a project functions is that there are shitloads of such slow nodes. An evolutionary algorithm is stochastic. This is especially true when searching along neutral networks [see Barnet reference below]. At this moment we are effectively performing a random walk along a neutral network in which we're facing the same kind of problem as above so slow islands will contribute proportionally to their processing speed. Ie. N islands of speed S will contribute as much as one island of speed N *S. Basically the more slow islands you add to the project, the more likely it is that one of them is going to come up with a fitter individual, or fit schema. Ie. even if some islands are not actually contributing they are POTENTIALLY contributing which is as much as you can say about fast ones.

Prok mentioned that a GA is useless with converged populations. This is old school Holland (1975) Schema theory. This has been rethought through and other evolutionary algorithms such as Goldbergs GA and Evolutionary Strategy (ES) have been used. The algorithm used in this project takes useful concepts from all these and specially from some rethinking about genetic algorithms performed at Sussex. I'll refer you mainly to SAGA theory explained in Harvey:92 . Another new interesting view of evolutionary algorithms is Barnet's Netcrawler algorithm .

Finally safemode is correct when saying that islands have an advantage if they have a faster interact and connect time. This only applies to the Knack ranking however, not to any other. To make it as fair as possible I could disllow more than 1 fitness improvement every N minutes for eg. I will add this to the list of things to do.

[safemode]

the stats may be the only thing affected visibly by people who choose to connect and/or interact very rapidly but the viability of the entire evolution of a circuit can be railroaded by it also. we can see very clearly from the block graph of the islands how rapidly some people have their interact and connect times. In less than 12 hours or so the circuit that srr5a has had spread to something like 25+ percent of the total islands in some way or form. interaction/connect has to be controlled by either an equation to set the ratio according to generation / speed or at least some lower limit.