Starcraft AI blog

Steamhammer 1.2 should play zerg a smidgeon stronger than 1.1.1, because of a few fixes. But I doubt the difference is noticeable. I didn’t fix anything important to strength.

After 2 days, Randomhammer’s rating has settled near 2050, a little above the average rating and 200 points below Steamhammer. The bot’s protoss and terran came out far weaker than its zerg. I’ve been looking at the games and thinking about why.

The building placement slowdown bug turns out to be critical for protoss. Randomhammer protoss has reached winning games against tough opponents like LetaBot and Killerbot by Marian Devecka, but tough opponents do not roll over and die, they fight stubbornly. Randomhammer fills up its main with widely-spaced buildings, the building placer slows to a crawl, and the bot busts the time limit and kicks out a loss.

Indecisive army movement is also critical for protoss. It’s harmful for zerg, but deadly for protoss, because the bulky protoss army spreads across the map, and one dragoon or another will happen to see something unimportant that distracts the attention of the whole army. The correct fix is to replace the targeting priority system, a job that needs a lot of detail work. I may try a cheaper fix instead.

Incompetence with terran seems harder to improve. Terran needs a lot of skills to play well, including adaptation skills. I did find the reason behind some strange play when going siege tanks, and fixed it.

Randomhammer beats Steamhammer on average, because Steamhammer’s openings are weaker versus random. When you don’t know what you’ll face, you have to play a general-purpose build. For example, when Randomhammer rolls zerg, it usually plays an optimized anti-zerg opening while Steamhammer is stuck with a tamer 9 pool or overpool. Steamhammer does not have a pro’s ability to play a general-purpose opening as sharply as a specialized one.

As a result, having Randomhammer in the pool of opponents decreases Steamhammer’s equilibrium elo, especially if people think the matchup is interesting and vote for it. Well, difficult opponents make good inspiration—all the better if I can supply the difficult opponent myself.

changes to plans for the next versions

Critical problems are a good excuse to rework my plans. Of course the real reason is that no plan survives contact with the calendar—it has been two days already, it’s time to change it up. The schedule must be rendered meaningless!

I decided to split versions again, so I can make the most important fixes up front and see the results sooner. The fixes now planned for 1.2.1 should delight protoss and leave terran and zerg smiling too. I don’t want to go too deeply into protoss and terran skills, though, because they’ll distract me from my main goals.

1. 1.2.1 - fix critical bugs, add critical skills
2. 1.2.2 - further skills
3. 1.2.3 - other lingering bugs