Starcraft AI blog

I uploaded Steamhammer 3.6 to SSCAIT. It turns out that Steamhammer terran needs a bug fix, so no Randomhammer for now. It will have to wait for my traditional post-AIIDE bugfix version.

As I mentioned earlier, I’ll post the change list after I’ve fixed up the blog. First things first.

Steamhammer is submitted for AIIDE 2023. You bring the futile hopes, I’ll bring the lurker spines.

It was kind of a scramble this year, though. Yesterday when I did what was supposed to be the final round of major tests, I turned up 3 critical bugs, 2 in new features and 1 left over from who knows when, hidden from view because the opponent model learned to avoid playing into it. As I stayed up late last night frantically debugging, a cricket got into the house—into the same room—and started talking loudly with its friends outside. Not entirely relaxing! But I got up early in the morning and finished fixing everything, and ran more tests, and packed it up, and ran the did-I-pack-it-up-correctly checks and tests, and now it’s on its way.

Last year, Steamhammer scored about 8% against BananaBrain, which finished first. BananaBrain has improved since then, but I expect to score substantially higher against it this year. Well, I could have missed further bugs, but there was no sign of it in what I had time for. My very last test game was against BananaBrain. It was hard-fought and went to late game, and zerg lost narrowly. Steamhammer managed to parasite every shuttle, and plague a lot of armies.

Next: Fixing the blog comments is my first priority. Then I’ll release Steamhammer and write up its changes.

Three items.

1. Steamhammer is coming along nicely. Today I completed and tested new infrastructure that will make it easy to solve one of Steamhammer’s biggest weaknesses, and bring other play improvements closer too. There’s not much time before the AIIDE submission deadline, but even so, expect harder fights.

2. Blog comments turn out to not be the only problem with the blog. I need to do some sysadmin stuff that I find tedious, and I don’t want to be distracted from working on Steamhammer. I’ll get around to it sometime in September for sure. Sorry. :-( I miss the good comments.

3. CoG has not posted results on its website, not even summary results. If and when they appear, I’ll do at least some of my usual analysis.

To fill the dead air while I try to fix the absurdly undiagnosable comment bug, here’s a game I was pleased with.

Steamhammer - BananaBrain on Python was played not long ago. (Here’s the replay.) BananaBrain held the initiative for most of the game, and broke into the zerg natural at one point, but never destroyed a completed zerg base. The game has plague, parasite, and broodling.

Steamhammer’s favorite strategy is to overrun you in the late game. Even strong opponents lose if they let the strategy play out.

AIIDE. Steamhammer will play in AIIDE again this year. I expect to bring a new version—it has been too long. Steamhammer’s infrastructure for some important new skills is almost ready, so if all goes as it should I will shatter your futile hopes with lurker spines!

I like the map rules this year: Five chosen from last year’s maps, five from a long list of varied maps. It’s a compromise between the guaranteed quality of a fixed map pool and CoG’s historical plan of selecting the maps from a list with inconsistent quality, some of which bots struggle on. The AIIDE list is large but sorted by quality, and the allowed maps have relatively few features that bots can’t cope with. Well, the list is long enough that there are several maps I don’t recognize.

I have often tested Steamhammer on random maps from a large collection, including utterly nonstandard Blizzard maps. It’s not good (yet) at taking advantage of map features, but it’s solid and consistent on standard-ish maps. It may be to my advantage.

CoG. The conference runs from 21-24 August, so results should be out soon. The roster includes one new bot, ProtossPleb, whose name and information does not look promising in itself. But we’ll see. The other participants are several familiar top finishers.

ASL Season 16 started this week. No great games so far, but I enjoyed Calm’s clever play against Rain in group B of the round of 24. The best games are often in the round of 8.

I tried turning comments back on, but it seems the feature grew a technical problem on the new webhost. Stand by for repairs.

I re-enabled Randomhammer on SSCAIT. It was disabled for the tournament. There are no updates to the code.

I disabled it on 19 December by uploading an empty zip. Strangely, Randomhammer’s game history shows 5 games played between 30 January and 4 February (today). Since there was no bot to run, the Randomhammer player did not do anything.

Steamhammer won all but 2 of its regames, shifting its score up minutely from 79-33 to 80-32. It is tied with WillyT, which means it’s not definite which opponent it will face. Depending on whatever rule is used to pair tied bots, it might face Hao Pan, PurpleWave, or Dragon. PurpleWave is a possible opponent only if Hao Pan wins its last regame, which is likely.

Halo by Hao Pan would be its ideal opponent. Steamhammer would stand a good chance to upset the higher ranked bot.

PurpleWave would almost certainly win.

Dragon is in between. Steamhammer’s score ratio versus Dragon on BASIL is around 2:3, so it has a theoretical chance of winning a match. But Steamhammer will do worse on SSCAIT because it has less data to go on. There would be a slight chance of an upset.

I may not have updated Steamhammer in 2022, but I am still following SSCAIT. I’m watching a lot of the games, and especially all Steamhammer games.

Krasi0 is participating, after sitting out recent tournaments. It has been missed.

I expect the top to be fought between Krasi0, Stardust, and BananaBrain. Krasi0 is the favorite to win the round robin phase. Monster has slipped a little as a few opponents have learned how to deal with it. PurpleWave put on a clinic for how to beat Monster as protoss, and so did Krasi0 as terran. Both won with one-base timing attacks at an early point when Monster’s build was weak. Monster needed better understanding of the opponent’s build to adapt. (Monster’s ZvZ build and adaptation are different, so zerg can’t simply copy the timing.)

PurpleWave is following its pattern of excellent preparation against key opponents it will face in the elimination phase, along with stray losses to weaker opponents. Will it be good enough to win the elimination phase and be the overall winner?

So far Steamhammer has been holding a position at the bottom of the top 16, with only slight wobbles. But the tournament is young. I expect it to finish about the same as last year. Only its top opponents have improved much, and they are all ranked higher anyway! And Steamhammer’s learning tends to make it robust against moderate updates to its opponents, at least once it has had time to adjust. But then, I have been over-optimistic in the past.

Steamhammer did not record all games, only the last 200. I’ll mention my preparation for different opponents, because it makes a big difference.

Steamhammer’s gas steal achieved nothing significant against any opponent. Maybe next year I’ll disable it.

In this game, Steamhammer hit on a surprising plan that exploited a weakness in Stardust. Even so, Stardust had tricks up its sleeve.

It is strategically correct for the stronger bot to play conservatively, taking no risks. Steamhammer happened across an opening which Stardust answers too conservatively: It proxied in Stardust’s natural. It started the hatchery almost immediately after scouting the location of the protoss base. Correct play is for Stardust to smash the proxy before it can be defended. Stardust scouted the proxy and immediately assumed without evidence that it was contained, so it played conservatively and did not try to break out. The yellow dot in Steamhammer’s natural is the scouting probe: Stardust had all the information it needed to conclude that the proxy was indefensible, but assumed that it was too risky to attack.

The production queues tell each bot’s strategy. Stardust made a robo to escape the containment by air. Also notice the Citadel of Adun. The two cannons next to the nexus are unnecessary; zerg does not have a lair, and protoss could have scouted it but did not bother to. Steamhammer’s opening build assumes blindly that the opponent will not attack—it is a build specialized for defeating one-base protoss players who build up and attack late. Steamhammer is making drones now in preparation for sunkens at the proxy and then a large army.

Both sides have attack +1 already. The citadel was for zealot speed. A shuttle can carry four zealots but only two dragoons. Stardust cleverly elevatored zealots from its main to the north of its base where they would not be seen, and ran them across the map. Steamhammer was ready anyway. Zerglings are about to hatch, and once they joined the hydras the zealots were afraid to engage and ran away. The zealots tried to retreat to the protoss main through the proxy, where sunkens and the zerg army slaughtered them.

After that zerg was in charge. Steamhammer immediately invaded the protoss base, while Stardust airlifted a probe out for a distant hidden base. There was more fighting, but Steamhammer is reliable about winning when this far ahead.

On Jade with its low main, it’s important to defend above the ramp if you can. Otherwise you don’t see what’s coming and you have to fight uphill. But as far as I can see, it didn’t affect this game. Stardust scouted the proxy and did not try to defend its ramp, except for one cannon.

Update: Steamhammer played a second game against Stardust, on the map Python. It went very much the same way. I was right that Jade’s low-ground main did not matter.

Second update, 10 July: Steamhammer has since won a bunch of games that went the same way. Here’s the first game that went differently, on Andromeda. It shows both the strength and the fragility of a strong proxy position.

Steamhammer-Krasi0P on Python (BASIL replay) was hilarious. Don’t miss the third zerg base in the upper left.

It’s easy to guess how the game went.

There was a somewhat similar game in March last year. In that game, Krasi0P put up only one containing cannon. In this game the protoss containment was effective for a long time.

Working at glacial speed, I pieced together a machine learning algorithm for use in Steamhammer. It passed basic tests today. When I’m sure that it works for real cases that I care about, I’ll release the code. It’s short.

I used a little-known idea and combined it with a standard algorithm employed in a slightly old-fashioned way. It took me a long time to put the thoughts together, but the result is that it fits in 374 lines of code, including blanks and comments, with no dependencies. (It will likely be longer when I release it.) It should be as accurate as a conventional (non-“deep”) neural network, but it learns faster and runs faster. It’s there in the literature, I just had to put the pieces together.

Hooking it into Steamhammer will take more code that the method itself. I’ve claimed before that the first use will be for an evaluation function, “Am I winning?” but that’s not strictly true. The first use will be to train the evaluation function! Once it has useful accuracy it will be put to work in opening selection, and it can be bootstrapped for all kinds of midgame strategy decisions.

Don’t hold your breath, I’m still working at glacial speed, and even global warming won’t change that right away.

I re-enabled Randomhammer on SSCAIT. There are no updates, I only turned it back on.

I predicted Steamhammer to finish at #11 in SSCAIT this year, and hoped it would do a little better. It finished tied for #12-13. On the one hand, it’s only a little lower than I expected. On the other, the difference in games from what I expected is glaring, to my eyes. When I made the prediction, I didn’t realize that Steamhammer’s saved learning data had been reset at some recent time. In the games I saw, Steamhammer had about 8 past games of data on each opponent. I did not imagine that Steamhammer might lose 2 games in a row to XIMP by Tomas Vajda, and 2 games in a row to WuliBot, and other losses to fixed-strategy opponents—it simply doesn’t happen when Steamhammer is trained up.

I estimate that if Steamhammer had won its “easy” games at the rate it does on BASIL, it would have finished at #10, with a chance of reaching #9. It would have been as I hoped.

Today’s finals round 1 match against Halo by Hao Pan was awful. Steamhammer scores over 60% versus Halo on BASIL. In the SSCAIT round robin it scored 2-0 using a ling flood strategy, which won when Halo opened its wall prematurely. In today’s match the ling flood failed, though it was close. Steamhammer didn’t have much experience to back up its next choices, and made poor ones.

Steamhammer’s next match is in the loser’s bracket against #13 McRaveZ. I think its odds are under 50%.

What happened in game 947, Steamhammer-Florian Richoux? It wasn’t the failure to connect that has disturbed other games. It looks like a related but different server failure.

Both bots recorded replays. Both replays are 3 seconds long. Florian Richoux (aka AIUR) recorded a replay where both bots sent workers to mine, end of game. Steamhammer recorded a replay where it sent drones to mine while Florian Richoux was idle as if it had not connected. The official result has Florian Richoux winning, and the game is not considered a crash.

I guess Steamhammer connected and then somehow lost its connection after it issued its mining orders and before Florian Richoux’s mining orders reached it? Or something?

So far Steamhammer has 4 games out of 34 played which were disturbed by apparent server failures. 2 are wins and 2 are losses. That’s about 12%, consistent with the estimated 14% overall rate from earlier on. The failures are adding noise and on average causing scores to shift toward 50%.

SSCAIT has only been underway for a short time. Results so far are very rough and will change. Even so, Steamhammer is scoring about as expected, currently 10-4 for #10. It has played more games than most bots. A good sign is that it has played more games against the top 16 than any other bot in the top 16, and still held its expected position.

A bad sign is that Steamhammer has two wins over opponents that did not start up: Halo by Hao Pan and Stardust. Stardust has 3 losses, all against opponents it should beat easily. None of the 3 has a replay recorded on Stardust’s side, so it must have failed to start all 3 games. If it’s the server’s fault, either the server bug has a bias or else Stardust is extremely unlucky. In a real game, Steamhammer has good odds against Hao Pan (better than 2:1), but virtually no chance against Stardust.

The ranking will change a lot before the end. So far, BetaStar and PurpleWave have perfect records with 7 and 6 games played respectively. BananaBrain, Monster, and Krasi0P follow with around 90%.