Our Genetic Program

[Cthulhu]

Well....acually I was going to write how I dont think your bot is going to make it to the top 100....only to find out that you already settled in only a few ELO-Points away...
Did I understand you correctly, that your entire code expect input/output is genetically engineered? Its an absolutely interesting thing to do - I just didnt think a useful bot would have come out of this in such a short time...actually I expected the to finish before you do...
I would have gone for a mixture of a handmade bot with only specific parts (like target evaluation) subject to genetic algorithms, but your bot proves you right. Looks like I have to update my code as well...somehow I dont like the idea of performing worse than "random-monkeylike-gibberish-code"-bot
Might start with my planet evaluation that currently depends on anything BUT the planets own growth rate...

[KonaEarth]

I understand you not wanting to release your primary code. Please though, if at all possible, let us see the generated code. I know it will probably be a mountain of mess. I don't care, I still think it will be fascinating. Just like the FPGA experiment (http://biology.kenyon.edu/slonc/bio3/AI ... _algo.html), I'm sure there will be some unexpected results.

[space.invaders]

[aerique]

[Innominate]

The only problem with "cheating" like that (besides some computational theorists tut-tutting over in academia-land) is that your derived parameters may not be optimal. Even this isn't a problem, unless they're close to optimal, in which case your programs will never evolve the optimal derivation function if there would be a big "valley" between them.

Say you have a program which at some point uses a derived parameter P, used in your most successful programs for some purpose. Assume that there is another possible derived parameter O which is "better" than P for that purpose (meaning, if O was used instead of P, the program would be more successful on average). If P would require a non-trivial number of mutations to reach O, and the intervening parameters are worse for that purpose than P, then no program will ever evolve to use O instead of P (or at least, it would be unlikely). On the other hand, there might be a route from some un-derived parameters to O which would lead to steady increases in performance. But since the initial stages of that route are eclipsed by P, nothing ever reaches O.

Imagine that it's foggy and you can't see far. When climbing a mountain, you decide to assume that you've reached the top when you get to the highest point that you can see. When climbing from the left, the first peak you reach would be the very top - O. When climbing from the right however, the left-hand side of P looks like the other side of the mountain, and so you stop at P.

Of course, that's a big set of ifs right there. Unless you chose really weird derived parameters, it shouldn't occur. And even if it did, the relative performance gain from getting to better bots sooner would probably dwarf the marginal utility of having the "best" function for a small portion of the bot's logic. I think you're doing exactly the right thing, which is proven by your bot's performance.

[space.invaders]

[space.invaders]

[space.invaders]

[space.invaders]

Our current status is somewhere between desperation and hope.

It is a disappointment that our last two daily submissions did not improve the ranking.
We still have hope that this might change due to a little observation.

The evolution of our bots playing capabilities goes in steps. Innovation breakthroughs, as we like to call it.

• It starts with finding out that sending fleets is generally better than not sending fleets. Big deal.
• Then, they find out that it is important to select good targets. It is generally better to send fleets to enemies and neutrals than just in circles within the own territory. Near planets, fast growing planets and weak planets also make better targets.
• Then, the first real eye-opener: planet sniping. An innovative way to get planets cheaply.
• Next logical step is to avoid having your own planets sniped.

That's where we are now in our evolution, and that's all we have seen in our test runs. Our strategies are not perfect in these areas, but they are doing fine, and they slowly keep improving over time. The source of our hope is that we started to see a new pattern for the first time. One you all know well:

• Building supply routes.

Seems like you can't make it into the top 100 anymore without a decent strategy to distribute ship within your own planets. Our current bots don't have a good supply route mechanism, but we are excited to see them "playing around" with that idea. Look at this game: http://ai-contest.com/visualizer.php?game_id=8651838. At turn 105, we see our right hand side planet sending ships to the center rather than directly to the enemy, and we see two small planets near the bottom sending ships to the bigger planet, which is closer to the enemy. Now, I understand that bocsimacko will not lose much sleep over that little hiccup in a third league match, that didn't even make any difference in that particular game. But the fact that our best bots start using that technique is a sign that they have successfully used it in their evolution, and if that is the case, I'm confident that they can improve it.

The next steps, which we will unfortunately not have time to see in our evolution run before Saturday night, would possibly be

• Find a way to play against good supply routes.
• Find solutions to all the little issues that we have not even identified as problem yet. That's where bocsimacko is already.

[Ice_Harley]

space.invaders, doesn't matter what and how your bot does, but reading your posts is great pleasure. Very funny. Thanks