Observation with Evolution

[rathaman]

Is it just me or does the evolution seem to stagnate once there are alot of "holes" in the topology? I know the current evolution is more involved, but it seems to be taking get stuck alot in one place. Could this be due to the new method of islands, or is the GA just "hitting a wall"?

[rshepard]

My guess is that we're "hitting a wall" - If you look at the listing of the islands, it looks like the top 380+ all are at the same state; so the "genetic material" is getting passed around in spite of the holes. We're just having a hard time making progress from this point.

[rathaman]

let me explain a little better... im no GA expert, but wouldn't all those 'holes' (ie dead islands), really hinder the evolution, since an active island is gion to be neighbour with a dead one. Now with there being a lot of dead islands, its giong to take alot longer for two active islands to communicate their best results to each other.

could someone let me know if this is what its doing, or if its designed differently?

[Stephen_B]

I think it will take a little longer for islands to pass their best individual along to each other, but in this case it doesn't matter since the best individuals are all alike.

I don't know how much diversity is present in the current island population (Miguel?). Ideally you'd like to keep your local maxima hanging around. By this I mean, even if you have this great individual that at the moment tops the charts in fitness, you don't want it to completely wipe out other good solutions, particularly if those other good solutions are different from your current one.

The reason you want the good-but-not-best individuals to hang around is that they contain information which might be used in a future better solution. You can quantify how different a solution is and use that information.

There's also a trade-off involved. The faster your population converges to a solution, the quicker it loses diversity.

[michaelgarvie]

I am planning to implement a new migration scheme whereby the destination for your next emigrating individual is chosen at random with a bias towards choosing closer ones.

For now, there is a small kludge which allows the fittest individual found so far to fall from the sky onto isolated islands.

[michaelgarvie]

Distant migration has been implemented. The maximum distance is 4 squares. Migration is guaranteed to find a partner at each client interaction. Probability of picking a partner is inversely proportional to its distance to your island.