Hacker News Reader

Comment by narsil | Hacker News Reader

The faq links for each bot are interesting. The author of the bot ranked first (Tscmoo) doesn't think the bot uses any interesting AI techniques or algorithms: https://dl.dropboxusercontent.com/u/23817376/Starcraft/AIIDE...

> My bot has a variety of different strategies that it can execute. The strategy to execute is chosen randomly using weights, and the weights are adjusted based on previous match results. It's pretty simple but I guess it could be called learning.

The top 3 bots are independent entries, and beat the bots using learning algorithms such as UCB1 from research institutions. The author of the bot ranked second does think that other bots that minimized hand-crafted solutions would be more interesting: https://dl.dropboxusercontent.com/u/23817376/Starcraft/AIIDE...

View on HN

The winner is very interesting, while the runner up uses the simplest possible strategy every single game. The runner up had a favourable record against the winner too, but lower win percentage overall. Thank goodness the more interesting one could win.

Basically, 4 pool beats all but one of the bots in the game over 50% of the time (often above 80-90%).

Developer set parametres are pretty common in industry "machine learning", the kind of thing Google NEtflix and Amazon use. Academic models aim to reduce the amount of human input, but most real world applications have no such motivation. This kind of AI can be seen as a developer parenting a machine.

I'd noticed too that the winner claimed to use just a variety of hand-rolled strategies, and wonder why that is. Has anyone had the time to dig into the code? Is this a case of well-writen "naive" code trumping theoretically-better but poorly-implemented algorithms, or are there properties of the game state-space which make e.g. Monte Carlo tree search perform poorly?

I think its just a case of this still being pretty difficult. So a decent handcrafted strategy will beat a computer generated one in a tournament.

Computers can't beat Go, either, not yet, though there's a lot more focus on this.

Speculation: the winning author is simply good at StarCraft and knew what to implement. His implementation was probably simple but did one thing very well, which beat more complicated/smart AI implementations from worse players.

Monte Carlo over the StarCraft state-space would be very hard to implement but I don't see any particular reason why it wouldn't work decently well.

The winner's bot had 16 different build orders it could play https://dl.dropboxusercontent.com/u/23817376/Starcraft/AIIDE...

Reply on news.ycombinator.com