[Melbourne-pm] AI Contest - some strategy

[Toby Corkindale]

On 11/11/10 17:44, Nathan Bailey wrote:
> My goal is to get perl into the top 100. Currently, the best Perl bot is
> ranked 175.
>
> Obviously I'd like to do that myself, but I'm also happy to see someone
> who is a better coder, mathematician or strategist than me get there, so
> here's some basic thoughts from me to help you get started, if you
> haven't already (and it's not hard to get started!)
>
> Given that this is a clearly the kind of AI problem that AI people find
> trivial to solve in LISP, I'm actually surprised that LISP doesn't make
> more of an appearance in the top 100, but the #1 bot is a LISP bot.
>
> Every game played for the rankings is recorded, and you can watch them,
> which gives you some insight into people's strategies - you can play
> back games turn-by-turn, seeing where they attack first, how many ships
> they send and which planets they consider most valuable. It's definitely
> worth doing this for the #1 bot. It's possible that you'll come up with
> an even better strategy that outthinks the current top ones, but the
> current ones have some common elements that are easily determined by
> watching.
>
> About 50% of the time, when I think I've improved my bot, I've actually
> made it worse. So before submitting it, I play it against itself. Toby's
> distribution includes 100 maps; I play my new bot against my old bot for
> all 100, and if my new bot doesn't win a majority of the games, I rework
> it!

I'm not convinced of that strategy, btw.
I was chatting to someone else in the competition who was trying out 
genetic methods of improving his bot automatically.
He could get it so that every new generation would find a version that 
was better than the previous.. but over time, actually became worse at 
beating other bots.

I've seen something similar too.. my current bot is much smarter than 
the early one, yet the early one's strategy was a bit more random, a bit 
more defensive, and actually does better against some specific types of 
enemy bots.

Actually, for all the extra analysis the current one does, it still does 
some dumb stuff at times :/

> I've been thinking about saving the maps each time I play over TCP; I
> suspect there are as many as 1000. Some of the maps have significantly
> different configurations (eg. home planets very close to each other at
> the start of the game) that are likely to affect your strategy.

I'm pretty sure you can download the TCP maps for that TCP server from 
somewhere.


> I've also been considering writing a merged bot, based on which types of
> maps each variant of my bot performs best at - or perhaps I should just
> work out why that is the case and incorporate it into my strategy :-)

The latter would be best!


> Although I've joined Toby in the top 10, I don't think my bot is that
> sophisticated, and I'm sure I'll be overtaken once y'all have some time
> to improve your bots :-)

What's your username?

I uploaded a new bot, and despite not getting many games given to him 
yet, my rank dropped first to 13, and then I uploaded another one and 
now I'm all the way down to 29! WTF?
I'm guessing this is a "feature" of the ranking system, and hopefully 
I'll get back up once I get a few more games played.

TOby