Bot Castro to reach first Havannah League

Timo Ewalds bot Castro has won
group 2.2
of the 9th Havannah Championships. So, Castro is allowed to start in the first division next time.
Congratulations to Timo!

Wow! One should rename it to 'master of sudden death'. Interesting to see whether or not it will survive in the first league.

Where can I get information about the parameters? Are the parameters changing? Are the parameters tuned by a human? Where can I get information about the machine it is executed on? About statistic information?

Anyting we have at mindsports is listed Here.

Very interesting indeed. Thanks for posting.

Impressive. Congrats!

Thanks! It's exciting to see Castro doing so well!

I haven't changed anything about Castro in quite a while. It is running on an Intel i5-2500K CPU @ 3.30GHz using 4 threads and 4gb of memory. It uses minutes per move, so 10 minutes per move in these tournaments. The parameters were hand tuned about a year ago but mainly based on extensive testing in self-play. You can read about the testing in my thesis: http://havannah.ewalds.ca/ . Overall Castro has over 3 million games logged as part of that testing. It's worth noting that none of that testing is against humans nor with such long time limits, so the testing is not entirely representative of these conditions, hence my little hand tuning. The parameters are similar but not quite the same as what was used against Christian. I'm not actually sure which is stronger, especially given the different time limits.

Is there any other information you'd like to know?

Thanks, I got some impression.

It is really painful to keep track of the ring threats that beast is generating. My first try was quite a fast loss.

Game #1520932

My fault. However, in the next game I tried to interact earlier and thus reduced the number of possible rings.

Game #1521096

I won also because it does not really try to achieve other goals like forming forks or bridges or prevent the opponent to do so.

In some sense this is not like humans play, on the other hand it is covered by the rules. Sometimes players like cutecat and also others act similar but for now I guess that there is no new insight except that one can reach league one by mastering unclear positions.

One of the things I tweaked when playing against Christian is how strongly to ignore rings. I can easily make it play towards the other goals by simply telling it that distant rings aren't a big threat, but then it becomes blind to them and becomes quite exploitable. I'm not quite sure where the optimal point is in that tradeoff. I'm fairly sure Castro does better on this metric than all the other bots, but there's definitely big room for improvement yet. There's a big section on this in my thesis. I can change it a bit if you'd like to try a less ring focused game.

Feel free to change the parameters. But maybe it would be good to keep them, just to have the same entity playing in league one that played in the lower leagues. Also I like the principle of least surprise :)

BTW: There is the possible attack of playing more than 36*6 invited games besides the tournament in order to make it running out of time. A trivial random bot could handle that. Would you like to share your s-cripts that interface to LG, please?

Well, I meant more as in you could play against it with weird parameters for one or two games, not for the tournament. For example I could make it completely ignore rings as a winning condition, but then you'd have to agree to never win by a ring. It'd be quite a different game, but also interesting.

True, there is the attack of just spamming it with games, but then again, you'd have to play them too. If you don't play well enough, the games will be quite short and it'd survive the attack.

The scripts are on my github account.

Btw, the script is here: https://github.com/tewalds/castro/blob/master/lg-castro.rb , with the full bot code being one level up.

Very interesting stuff :)

In that case, Ingo Althofer sent me this mail:

"Now two reports on your Havannah match were published in the ICGA Journal.
You can find the table of content of that Journal issue in the appendix."

Attached was this jpg:

So.. now the league should start soon. Will be interesting to see how it does in league one.

Is there somewhere were we can read the report from Tony?

The report from Tony is similar to what was posted on G+.

Ingo's report on the other hand is not particularly good in my opinion. It sounds like Havannah is popular mainly because of Ingo himself. It completely fails to mention that I wrote Castro as part of a class by Martin Mueller (author of Fuego) where 7 Havannah programs were written in parallel and had certainly never heard of Ingo before writing. From what I remember from Marcin, he chose to work on Havannah independently as well.

He also claims I wanted to capture the win and was disappointed. I'm happy to have gotten a win, but don't really care who got the first win. I'm certainly happy that Marcin got a couple wins as well.

He claims we didn't talk much, but that's not true either. We ended up collaborating a fair amount and were working on a script to have the two bots play cooperatively. Unfortunately time constraints stopped that from actually happening.

He also claims we didn't want to write a paper. It'd have been nice, but again, time constraints. I'm also not sure what I could add above and beyond my thesis and source code though, which is much longer and more detailed than any paper I could possibly submit.