Starcraft AI blog

Change of plans. Today and tomorrow I’ll download my thoughts about Steamhammer 1.0 before I dive into 1.1 and start to forget which feature came when. After that I’ll return to tournament coverage. Keeping up with real time? I may have heard of that.

Steamhammer’s web page is updated. Get the 1.0 source and binary releases there. Downloading from SSCAIT is another way to get the binary. Don’t miss the config file, the bot needs it to run.

new skills

The configuration file lets you specify more possibilities as part of the opening build orders. One of my goals is to be able to play every opening, and I’m making progress.

• scout on command: always, or only if necessary, or only to find the enemy
• build a macro hatchery, or mineral only base, or a gas base, or a “hidden” base
• stop and restart gas gathering, or gather a specified amount of gas then stop
• build static defense in the main or natural

For example, you can code “go scout if needed” to send out a worker scout, right then, only if the enemy base location is not already known. Or “go gas until 100” to get 100 gas for zergling speed and then put the drones back on minerals. Or “hatchery @ min only” to start a new base that does not need to have gas. These are items that can go into the production queue to be executed when they come up; the strategy boss uses the same underlying code to communicate its dynamic plans to the production manager.

New openings use all these features. The old openings in Steamhammer have also been reworked extensively. Even when it looks like it’s doing the same thing as the previous version, there are differences. Some openings have fancy optimizations, like stopping and restarting gas more than once to smooth production. It’s easy to edit the config file, so it’s no special effort to try out tricks like that.

The strategy boss copes with emergencies like having no drones and has some basic reactive skills, like queueing up scourge when corsairs are sighted. Overall Steamhammer is much more robust when things go wrong; there are still bugs that can freeze it, but only on cool days.

The strategy boss also improves macro. Steamhammer still sometimes runs its minerals up into the thousands, but it recovers reasonably soon, and in many cases macro is fine from the beginning of the game to the end. Far from perfect, still much better.

random openings

Steamhammer 0.2’s ZvT configuration looks like this. It plays one fixed opening.

		"ZvT" : "ZvT_12Pool",

Steamhammer 1.0’s looks like this.

		"ZvT" :
			{ "StrategyMix" : [
				{ "Weight" :  3, "Strategy" : "4PoolHard" },
				{ "Weight" :  2, "Strategy" : "4PoolSoft" },
				{ "Weight" :  5, "Strategy" : "5Pool" },
				{ "Weight" : 10, "Strategy" : "9PoolSpeed" },
				{ "Weight" : 25, "Strategy" : "ZvT_12Pool" },
				{ "Weight" : 35, "Strategy" : "ZvT_13Pool" },
				{ "Weight" : 10, "Strategy" : "ZvT_2HatchMuta" },
				{ "Weight" : 10, "Strategy" : "ZvT_3HatchMuta" }
			]},

The bot chooses randomly among 8 openings, 4 zergling openings and 4 mutalisk openings. I haven’t worked seriously on lurker skills yet. The weights can be any positive integers, but I chose weights that add up to 100 so that I can interpret them as percentages. The 4 and 5 pool openings are slight variations of the same thing and happen 10% of the time in total. (With better scouting skills and followup, Steamhammer plays these fast rushes more strongly than UAlbertaBot.) The 9 pool overruns surprisingly many terrans and happens another 10% of the time. The 12 pool and 13 pool are similar muta rushes and together happen 60% of the time. Most people won’t notice the difference between the similar openings, but many details vary. The 2 hatch and 3 hatch muta openings are actual genuine honest-to-Betsy true-blue manufacturer-guaranteed mainstream builds with stronger economy, but Steamhammer is not yet adaptive enough to play them safely, so together they happen only 20% of the time. I’m curious to see how well they do in practice—hmm, I predict results will vary!

Steamhammer also chooses randomly in the other matchups, including versus random. You have to scout it to know what it is doing.

other stuff

The configuration file includes the version number: Steamhammer_1.0.json. That way I can conveniently run more than one version from the same directory.

I blogged about the fixes for original UAlbertaBot bugs. There were 10 in total, and only a couple were already in version 0.2. Stream viewers may not notice, but Steamhammer makes slightly fewer absurd blunders.

There are many small improvements to tactics and targeting, which together make a noticeable difference. The bunker obsession is reduced but still claims its toll. The bot is less eager to send out drones for emergency defense, but doesn’t run them away when they are under attack (“keep working, you cowards!”). And Steamhammer now has basic overlord vision skills thanks to hints from AIL, so it makes more informed engagement decisions and can fight cloaked units like dark templar.

And that’s about it. All other changes are minor.

future plans

Alongside further work on adaptivity in the strategy boss, next I want to improve the defense against worker harassment. Aggressive harassers like Iron and Bereaver often kill 2 or more drones, a grave setback (very like being set back in the grave). After that, I want to finally fix the mutalisk tactics and put those terrans in their place. When the mutas meet a strongpoint they should not call a dance party, they should turn aside (“you’re no fun!”) and fly around it. After comparing ideas and thinking ahead to how I want to integrate it someday with machine learning, I’m leaning toward Berkeley Overmind-style convex hull analysis rather than Overkill-style pathing analysis or ZerGreenBot-style reactive movement. But we’ll see.

One thing at a time. Mutalisk skills first, then lurker skills, then drop skills. Once I have all those, opponents will need real smarts to foresee what is going to happen to them.

Tomorrow: Some of Steamhammer’s opponent-specific openings.