Starcraft AI blog

In 2020, I analyzed McRave’s three-part strategy representation. This version of McRave is the CoG 2022 version, and I think details have changed, but the basic structure is the same as in 2020.

With its deadly mutalisks, McRave apparently does not bother to learn what to play against zerg opponents. In ZvZ, it follows the same 9 pool into one hatch muta strategy every game. It went mutas in most games against terran and protoss, too. But it’s still interesting to see what it recognized in its opponents play.

Microwave is configured to save only the most recent 200 game records. There were 290 rounds, so where all games were played, it dropped data for 90 games. (Steamhammer is configured the same way, though the data it stores is quite different.)

Mutalisks were good—relatively speaking—even though this version of BananaBrain goes heavy on the corsairs. Even the slow-starting 3 hatch muta was more successful than most builds. It’s curious that the highest win rate came when Microwave was unable to figure out what opening BananaBrain was playing. I doubt the cost of scouting is that high. I’m guessing it’s more likely that scouting failed more often when starting positions were distant. That gave Microwave more time to prepare.

But Stardust was (a little) vulnerable to zerglings. Again, the highest win rate came when Microwave did not know what Stardust was doing.

PurpleWave just fell over to Microwave’s zerglings. But even here, an Unknown enemy build was good news for Microwave.

Steamhammer was also successful with zerglings (I’ll post its data when I get to it), but nothing like this successful. I expect that, true to its name, Microwave has superior zergling micro. Though maybe it’s just more aggressive?

A lot of openings are roughly equal. I suspect that some of the ones that scored 0 would have also been roughly equal if they’d been tried more times.

So: Good with zerglings, not so good with mutalisks. Microwave knows how to defend its bases with two spore colonies each, in the standard pro way, but apparently that’s not enough to survive against McRave.

The two top builds are 5 hatcheries before spawning pool. Compare Dragon’s table for Microwave, which says that opening with mass vultures was best. The vultures must have often come before Microwave could defend itself.

Once the opening build is over, Dragon adapts to the opponent’s unit mix and many of its games start to look similar. Nevertheless, the opening makes a big difference in how the game comes out.

The “dirty worker rush” is a late rush with a lot of SCVs. You have to be ready for it, but if you are, it’s easy to beat.

opening	total	#1 bananab	#2 stardus	#3 purplew	#4 steamha	#5 mcrave	#6 microwa	#7 dragon	#9 infeste
total	-  31%	5-279  2%	25-263  9%	53-236  18%	9-274  3%	183-102  64%	112-170  40%	48-236  17%	263-26  91%
4RaxMarines	8-117 6%	0-19 0%	5-33 13%	0-11 0%	0-22 0%	-	3-16 16%	0-16 0%	-
MarineRush	181-240 43%	2-25 7%	1-18 5%	14-54 21%	1-28 3%	65-30 68%	18-45 29%	3-29 9%	77-11 88%
TankPush	1-89 1%	0-18 0%	1-16 6%	0-11 0%	0-21 0%	-	0-7 0%	0-16 0%	-
VultureRush	1-90 1%	0-18 0%	0-12 0%	0-11 0%	0-21 0%	-	0-7 0%	1-21 5%	-
DTRush	24-177 12%	1-40 2%	0-25 0%	3-41 7%	0-18 0%	-	-	20-53 27%	-
DragoonRush	2-119 2%	0-32 0%	2-39 5%	0-22 0%	0-18 0%	-	-	0-8 0%	-
ZealotRush	229-236 49%	0-32 0%	2-39 5%	0-22 0%	8-54 13%	62-46 57%	59-23 72%	3-18 14%	95-2 98%
2HatchHydra	44-143 24%	1-27 4%	0-10 0%	32-44 42%	0-23 0%	-	2-10 17%	9-29 24%	-
3HatchMuta	1-73 1%	0-21 0%	0-10 0%	0-4 0%	0-23 0%	-	0-5 0%	1-10 9%	-
3HatchScourge	1-73 1%	0-20 0%	1-14 7%	0-4 0%	0-23 0%	-	0-5 0%	0-7 0%	-
ZerglingRush	206-229 47%	1-27 4%	13-47 22%	4-12 25%	0-23 0%	56-26 68%	30-52 37%	11-29 28%	91-13 88%

UAlbertaBot’s primary builds are MarineRush, ZealotRush (9-9 gates), and ZerglingRush (5 pool). Against #5 McRave and #9 InfestedArtosis, it stuck with them. Against other opponents it experimented with alternatives to try to escape its losses. It found 4RaxMarines (barracks at 4 supply) against #2 Stardust, 2HatchHydra against #3 PurpleWave, and DTRush against #7 Dragon.

InfestedArtosis learns three things independently of each other: Opening build, unit mix, and whether to build a defensive sunken under certain circumstances. For some reason, every opponent has data for exactly 46 decisions of whether to build the defense. I didn’t look into why.

It enables different options against each race, so I made separate tables. It didn’t seem worthwhile to write code to extract the small amount of data from 8 short files (especially the clumsy way I first thought of doing it), so I made the tables by hand. There may be minor errors.

The tables simply repeat what’s in the learning files. The first number is wins, the second is losses.

protoss

item	BananaBrain		Stardust		PurpleWave
OverPool	0	73	0	73	14	54
NinePoolSpeed	0	73	0	73	1	33
FourPool	0	70	0	70	35	43
TwelveHatch	0	73	0	72	31	76
LingFlood	0	73	0	73	77	115
Mutalisk	0	72	0	72	3	33
Default	0	72	0	72	1	30
Hydra	0	72	0	71	0	28
sunken	0	46	0	46	13	33

InfestedArtosis seems to have lost information about one game versus #2 Stardust, though the official results say that all 290 games completed and that InfestedArtosis had zero crashes.

#3 PurpleWave was vulnerable to zerglings, especially with 4 pool but also after the economical 12 hatchery.

zerg

item	Steamhammer		McRave		Microwave
OverPool	0	55	1	59	3	58
TwelvePool	0	54	0	56	1	52
NinePoolSpeed	3	64	1	56	7	66
TwelveHatch	1	58	0	56	0	49
NineHatch	0	54	0	55	0	49
LingFlood	1	137	2	148	4	132
Mutalisk	3	148	0	137	7	142
sunken	1	45	0	46	2	44

terran

item	Dragon
TwelvePool	15	53
NinePoolSpeed	13	52
FourPool	0	30
TwelveHatch	35	81
LingFlood	1	33
Mutalisk	10	45
Default	38	84
Hydra	14	54
sunken	8	38

Rushes did not work against the slow-moving #7 Dragon, but countering with slow zerg play was fairly effective. TwelveHatch 35/81 is 43%, while TwelvePool 15/53 is only 28%.

random

item	UAlbertaBot
OverPool	1	61
NinePoolSpeed	0	58
FourPool	18	65
FivePool	6	77
LingFlood	21	100
Mutalisk	3	59
Default	1	53
Hydra	0	49
sunken	2	44

I’m surprised that 4 pool worked as well as it did against UAlbertaBot. UAlbertaBot’s protoss and zerg builds should both counter it.

No big analysis, but I do want to point out a couple features of the cumulative results over time.

#1 Stardust, #3 PurpleWave, and #8 UAlbertaBot largely declined over the tournament, or declined and leveled out. They were outlearned by the competition. Steamhammer is slow to learn but keeps learning for a long time, and its curve is fairly flat.

The two bots with the most impressive learning curves are #7 Dragon, a carryover from last year, and #9 InfestedArtosis, the tail ender. Dragon improved between round 25 and 100. Infested Artosis improved after round 100. Between round 100 and the end of the tournament, InfestedArtosis increased its score by 3.1% in absolute terms, more than any other bot improved after round 100. No other bot comes close over the same period. The weakest bot outdid all others in continuing to play better over the entire long tournament.

InfestedArtosis has a simple learning method. It independently learns three things, each from a short menu of choices: What opening build to play, what unit mix to follow up with, and whether to build static defense. Imagine how well it might learn if it, say, approximated the joint distribution and chose based on that!

The first table is the same as yesterday’s, for reference. Except I didn’t mark old/new maps.

#	bot	overall	Destin	MatchP	Polari	Longin	NeoSyl	Arcadi	Coloss	Empire	NeoGro	Roadki
1	bananabrain	84.54%	91%	81%	81%	87%	87%	87%	84%	81%	83%	84%
2	stardust	83.43%	71%	87%	90%	89%	84%	84%	86%	74%	84%	85%
3	purplewave	57.70%	65%	59%	53%	44%	58%	60%	61%	61%	59%	57%
4	steamhammer	50.11%	53%	51%	54%	54%	47%	46%	53%	48%	48%	46%
5	mcrave	49.81%	53%	44%	35%	50%	44%	58%	45%	60%	61%	47%
6	microwave	43.44%	40%	47%	49%	41%	45%	33%	49%	51%	33%	46%
7	dragon	42.28%	44%	45%	43%	43%	46%	44%	32%	38%	44%	44%
8	ualbertabot	30.45%	22%	23%	32%	35%	34%	31%	36%	28%	29%	33%
9	infestedartosis	8.28%	10%	12%	12%	7%	5%	8%	4%	9%	8%	6%

One table for each bot, showing results on each map against each opponent. With 10 maps and 290 rounds, each cell shows the results for 29 games (some slightly fewer). It’s not a large number, there is plenty of statistical noise.

#	bananabrain	overall	Destin	MatchP	Polari	Longin	NeoSyl	Arcadi	Coloss	Empire	NeoGro	Roadki
2	stardust	48%	86%	31%	14%	45%	52%	69%	17%	72%	48%	45%
3	purplewave	58%	62%	45%	62%	69%	62%	57%	66%	52%	55%	48%
4	steamhammer	97%	100%	93%	100%	93%	100%	97%	100%	100%	97%	86%
5	mcrave	86%	97%	93%	79%	100%	97%	83%	100%	48%	66%	97%
6	microwave	95%	97%	100%	100%	97%	83%	100%	93%	86%	97%	100%
7	dragon	94%	93%	90%	97%	97%	100%	86%	93%	93%	100%	97%
8	ualbertabot	99%	97%	100%	100%	93%	100%	100%	100%	96%	100%	100%
9	infestedartosis	100%	100%	100%	100%	100%	100%	100%	100%	100%	100%	100%
	overall	84.54%	91%	81%	81%	87%	87%	87%	84%	81%	83%	84%

#5 McRave’s comparatively good results versus #1 BananaBrain are due to only a few maps. Interesting! For the most part, BananaBrain is consistent. It seems to have few weaknesses that other bots are able to exploit.

#	stardust	overall	Destin	MatchP	Polari	Longin	NeoSyl	Arcadi	Coloss	Empire	NeoGro	Roadki
1	bananabrain	52%	14%	69%	86%	55%	48%	31%	83%	28%	52%	55%
3	purplewave	55%	0%	55%	76%	79%	48%	76%	79%	21%	59%	55%
4	steamhammer	92%	100%	93%	90%	97%	100%	90%	72%	90%	97%	97%
5	mcrave	95%	86%	97%	100%	100%	100%	93%	90%	93%	93%	97%
6	microwave	90%	72%	93%	97%	97%	97%	93%	90%	66%	97%	97%
7	dragon	92%	97%	93%	72%	90%	97%	93%	93%	100%	93%	97%
8	ualbertabot	91%	100%	93%	97%	97%	86%	97%	79%	97%	86%	83%
9	infestedartosis	100%	100%	100%	100%	100%	100%	100%	100%	100%	100%	100%
	overall	83.43%	71%	87%	90%	89%	84%	84%	86%	74%	84%	85%

#2 Stardust is very inconsistent against its fellow top protoss. These are the results of a brittle bot, strong but not resilient. On Destination, I expect it’s because the twin bridges to the natural bases are easy to defend, so that Stardust’s attack fails. But Empire of the Sun and the other troublesome maps aren’t like that. I can’t guess the reason there.

#	purplewave	overall	Destin	MatchP	Polari	Longin	NeoSyl	Arcadi	Coloss	Empire	NeoGro	Roadki
1	bananabrain	42%	38%	55%	38%	31%	38%	43%	34%	48%	45%	52%
2	stardust	45%	100%	45%	24%	21%	52%	24%	21%	79%	41%	45%
4	steamhammer	84%	83%	97%	97%	59%	72%	100%	90%	79%	86%	79%
5	mcrave	60%	45%	79%	66%	52%	79%	38%	90%	48%	31%	69%
6	microwave	29%	62%	7%	41%	3%	17%	76%	3%	3%	72%	3%
7	dragon	49%	45%	48%	38%	41%	41%	31%	79%	71%	48%	52%
8	ualbertabot	82%	83%	86%	76%	76%	79%	79%	83%	90%	86%	79%
9	infestedartosis	70%	62%	55%	48%	72%	86%	86%	90%	66%	59%	79%
	overall	57.70%	65%	59%	53%	44%	58%	60%	61%	61%	59%	57%

And #3 PurpleWave is inconsistent against all opponents other than #8 UAlbertaBot—even the tail-ender #9 InfestedArtosis. Which is curious, because UAlbertaBot plays random and is hard to be consistent against!

#	steamhammer	overall	Destin	MatchP	Polari	Longin	NeoSyl	Arcadi	Coloss	Empire	NeoGro	Roadki
1	bananabrain	3%	0%	7%	0%	7%	0%	3%	0%	0%	3%	14%
2	stardust	8%	0%	7%	10%	3%	0%	10%	28%	10%	3%	3%
3	purplewave	16%	17%	3%	3%	41%	28%	0%	10%	21%	14%	21%
5	mcrave	50%	72%	79%	69%	52%	38%	24%	48%	45%	38%	31%
6	microwave	62%	69%	41%	72%	79%	45%	83%	66%	48%	66%	55%
7	dragon	67%	72%	76%	90%	55%	62%	52%	76%	66%	62%	55%
8	ualbertabot	97%	93%	97%	93%	97%	100%	97%	100%	100%	100%	93%
9	infestedartosis	99%	100%	97%	97%	100%	100%	97%	100%	97%	100%	100%
	overall	50.11%	53%	51%	54%	54%	47%	46%	53%	48%	48%	46%

This year’s improvements to #4 Steamhammer are particularly helpful against #8 UAlbertaBot, because of Steamhammer’s turtle strategy for surviving whichever rush UAlbertaBot ends up playing. The consistently excellent results against UAlbertaBot were key to squeaking past #5 McRave for the #4 spot.

#	mcrave	overall	Destin	MatchP	Polari	Longin	NeoSyl	Arcadi	Coloss	Empire	NeoGro	Roadki
1	bananabrain	14%	3%	7%	21%	0%	3%	17%	0%	52%	34%	3%
2	stardust	5%	14%	3%	0%	0%	0%	7%	10%	7%	7%	3%
3	purplewave	40%	55%	21%	34%	48%	21%	62%	10%	52%	69%	31%
4	steamhammer	50%	28%	21%	31%	48%	62%	76%	52%	55%	62%	69%
6	microwave	89%	100%	97%	7%	100%	100%	93%	97%	100%	97%	97%
7	dragon	65%	52%	48%	76%	69%	45%	76%	66%	76%	79%	62%
8	ualbertabot	35%	72%	55%	14%	34%	24%	31%	28%	38%	41%	14%
9	infestedartosis	99%	100%	100%	93%	100%	100%	100%	100%	100%	100%	100%
	overall	49.81%	53%	44%	35%	50%	44%	58%	45%	60%	61%	47%

Wow, look at #5 McRave vs #6 Microwave! Nearly all wins, except on Polaris Rhapsody where it was nearly all losses. The reason for it must be very specific.

#	microwave	overall	Destin	MatchP	Polari	Longin	NeoSyl	Arcadi	Coloss	Empire	NeoGro	Roadki
1	bananabrain	5%	3%	0%	0%	3%	17%	0%	7%	14%	3%	0%
2	stardust	10%	28%	7%	3%	3%	3%	7%	10%	34%	3%	3%
3	purplewave	71%	38%	93%	59%	97%	83%	24%	97%	97%	28%	97%
4	steamhammer	38%	31%	59%	28%	21%	55%	17%	34%	52%	34%	45%
5	mcrave	11%	0%	3%	93%	0%	0%	7%	3%	0%	3%	3%
7	dragon	56%	62%	55%	48%	59%	52%	62%	76%	52%	34%	55%
8	ualbertabot	61%	62%	66%	67%	41%	48%	55%	66%	62%	69%	72%
9	infestedartosis	96%	93%	97%	97%	100%	100%	93%	100%	97%	90%	96%
	overall	43.44%	40%	47%	49%	41%	45%	33%	49%	51%	33%	46%

The overall results make it look like #6 Microwave somehow has #3 PurpleWave’s number, but here we see it depends on the map. It is presumably part of PurpleWave’s general inconsistency, but we’ll have to look at the replays to see the specific reasons.

#	dragon	overall	Destin	MatchP	Polari	Longin	NeoSyl	Arcadi	Coloss	Empire	NeoGro	Roadki
1	bananabrain	6%	7%	10%	3%	3%	0%	14%	7%	7%	0%	3%
2	stardust	8%	3%	7%	28%	10%	3%	7%	7%	0%	7%	3%
3	purplewave	51%	55%	52%	62%	59%	59%	69%	21%	29%	52%	48%
4	steamhammer	33%	28%	24%	10%	45%	38%	48%	24%	34%	38%	45%
5	mcrave	35%	48%	52%	24%	31%	55%	24%	34%	24%	21%	38%
6	microwave	44%	38%	45%	52%	41%	48%	38%	24%	48%	66%	45%
8	ualbertabot	83%	86%	93%	86%	83%	90%	86%	59%	83%	79%	86%
9	infestedartosis	78%	90%	76%	79%	69%	76%	62%	76%	76%	93%	86%
	overall	42.28%	44%	45%	43%	43%	46%	44%	32%	38%	44%	44%

#	ualbertabot	overall	Destin	MatchP	Polari	Longin	NeoSyl	Arcadi	Coloss	Empire	NeoGro	Roadki
1	bananabrain	1%	3%	0%	0%	7%	0%	0%	0%	4%	0%	0%
2	stardust	9%	0%	7%	3%	3%	14%	3%	21%	3%	14%	17%
3	purplewave	18%	17%	14%	24%	24%	21%	21%	17%	10%	14%	21%
4	steamhammer	3%	7%	3%	7%	3%	0%	3%	0%	0%	0%	7%
5	mcrave	65%	28%	45%	86%	66%	76%	69%	72%	62%	59%	86%
6	microwave	39%	38%	34%	33%	59%	52%	45%	34%	38%	31%	28%
7	dragon	17%	14%	7%	14%	17%	10%	14%	41%	17%	21%	14%
9	infestedartosis	91%	72%	76%	90%	100%	97%	97%	100%	90%	97%	93%
	overall	30.45%	22%	23%	32%	35%	34%	31%	36%	28%	29%	33%

#	infestedartosis	overall	Destin	MatchP	Polari	Longin	NeoSyl	Arcadi	Coloss	Empire	NeoGro	Roadki
1	bananabrain	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%
2	stardust	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%
3	purplewave	30%	38%	45%	52%	28%	14%	14%	10%	34%	41%	21%
4	steamhammer	1%	0%	3%	3%	0%	0%	3%	0%	3%	0%	0%
5	mcrave	1%	0%	0%	7%	0%	0%	0%	0%	0%	0%	0%
6	microwave	4%	7%	3%	3%	0%	0%	7%	0%	3%	10%	4%
7	dragon	22%	10%	24%	21%	31%	24%	38%	24%	24%	7%	14%
8	ualbertabot	9%	28%	24%	10%	0%	3%	3%	0%	10%	3%	7%
	overall	8.28%	10%	12%	12%	7%	5%	8%	4%	9%	8%	6%

I like it that even the tail-end bot has an upset against the strong #3 PurpleWave on one map.

Five maps were familiar ones used in past AIIDE tournaments, and five were chosen randomly from a large pool.

carryover maps	new maps
(2) Destination	(2) Match Point
(2) Polaris Rhapsody	(3) Neo Sylphid
(3) Longinus	(4) Arcadia II
(4) Empire of the Sun	(4) Neo Ground Zero
(4) Roadkill	(4) Colosseum II

That makes 3 2-player maps, 2 3-player maps, and 5 4-player maps. It’s a good mix.

Each cell in the two tables represents 232 games, or occasionally 1 or 2 fewer if not all games completed successfully. I marked the carryover maps with dark headers, so you can look for patterns between familiar and new maps.

#	bot	overall	Destin	MatchP	Polari	Longin	NeoSyl	Arcadi	Coloss	Empire	NeoGro	Roadki
1	bananabrain	84.54%	91%	81%	81%	87%	87%	87%	84%	81%	83%	84%
2	stardust	83.43%	71%	87%	90%	89%	84%	84%	86%	74%	84%	85%
3	purplewave	57.70%	65%	59%	53%	44%	58%	60%	61%	61%	59%	57%
4	steamhammer	50.11%	53%	51%	54%	54%	47%	46%	53%	48%	48%	46%
5	mcrave	49.81%	53%	44%	35%	50%	44%	58%	45%	60%	61%	47%
6	microwave	43.44%	40%	47%	49%	41%	45%	33%	49%	51%	33%	46%
7	dragon	42.28%	44%	45%	43%	43%	46%	44%	32%	38%	44%	44%
8	ualbertabot	30.45%	22%	23%	32%	35%	34%	31%	36%	28%	29%	33%
9	infestedartosis	8.28%	10%	12%	12%	7%	5%	8%	4%	9%	8%	6%

This table is the same as the map table in the official results except for the presentation. #5 McRave and #6 Microwave have clear likes and dislikes among the maps. The other mid-ranked zerg, #4 Steamhammer, shows more even results.

#	bot	overall	Destin	MatchP	Polari	Longin	NeoSyl	Arcadi	Coloss	Empire	NeoGro	Roadki
1	bananabrain	10:47 14:37	10:55 13:18	11:12 14:42	10:52 15:58	10:34 15:32	10:43 14:00	10:53 15:43	9:58 14:55	10:38 14:55	10:54 13:46	10:56 12:44
2	stardust	10:43 16:51	10:16 17:34	11:07 18:25	10:48 21:00	10:27 17:22	10:54 18:08	10:58 15:44	11:06 14:51	10:12 14:40	10:31 18:22	10:45 15:19
3	purplewave	10:26 12:02	11:02 12:12	10:56 13:32	10:21 12:38	10:22 11:05	10:16 12:02	10:31 12:43	10:11 9:53	10:24 11:46	10:13 13:18	10:42 11:23
4	steamhammer	7:02 10:45	6:35 10:53	6:54 11:14	6:30 10:42	8:05 10:12	6:11 11:08	9:05 10:36	8:23 11:19	5:52 10:16	5:50 10:31	9:50 10:31
5	mcrave	9:15 11:27	9:21 11:56	8:52 12:31	12:37 11:07	8:27 11:25	8:38 11:13	9:20 12:11	8:43 9:55	10:39 11:46	11:58 12:21	8:49 11:12
6	microwave	7:03 10:05	6:29 10:09	7:52 10:45	6:34 10:40	7:05 9:47	7:19 9:47	6:26 9:58	7:04 10:14	7:01 9:28	6:09 10:03	7:02 10:43
7	dragon	15:45 12:05	15:41 12:04	16:38 11:53	16:08 11:36	16:04 12:49	15:54 12:07	16:03 14:31	14:11 11:10	14:35 11:42	16:02 11:47	14:42 12:39
8	ualbertabot	6:54 10:08	8:19 9:27	7:32 10:42	7:11 9:50	6:25 9:59	6:48 9:32	7:05 10:31	6:32 10:42	6:25 10:15	6:42 9:47	6:34 10:06
9	infestedartosis	9:52 6:31	9:20 7:50	10:09 7:36	8:47 7:02	11:55 5:51	10:36 6:08	12:15 6:26	9:51 6:33	12:31 6:12	9:23 6:10	12:06 6:18
	overall	10:18	10:12	10:56	10:14	10:03	10:00	10:38	10:04	10:08	10:18	10:22

This is like yesterday’s game duration table, but by map instead of by opponent. The top number in each cell is the median duration of winning games; the bottom number is losing games.

On maps where #2 Stardust won the most quickly—it didn’t win as much. On Destination, I imagine it’s because the twin bridges to the natural bases are so easy to defend; if it doesn’t break through quickly, then it may struggle against its stronger opponents. That explanation doesn’t work for Empire of the Sun, though. #7 Dragon has a similar pattern. Compare #3 PurpleWave: On Longinus where it struggled, it lost more quickly.

#4 Steamhammer takes the map into account in its learning, which may make its performance more even between maps, and its game durations less even. Possibly.

This crosstable exactly matches the official results. I had to make two adjustments compared to last year to get it to match:

1. In some failure conditions, the game is recorded anyway, provided it started in the first place. If the game did not start, last year its duration was recorded as 0 frames in detailed_results.txt. This year it is recorded as -1. The HTML detailed results still gives the -1 as duration 00:00:00.

2. Game ID 979, between UAlbertaBot and BananaBrain, is recorded in detailed_results.txt with finishing status NORMAL. But it’s obviously not normal; it looks like BananaBrain’s Starcraft instance crashed. There is no replay saved for BananaBrain, and no final score is recorded either. The official results correctly recognize that the game did not finish, and the HTML detailed results give the finishing status NO_REPORT, which means the game is not counted. But it’s still NORMAL in detailed_results.txt, so I coded it as a one-game special case for this analysis.

#	bot	overall	bana	star	purp	stea	mcra	micr	drag	ualb	infe
1	bananabrain	84.54%		48%	58%	97%	86%	95%	94%	99%	100%
2	stardust	83.43%	52%		55%	92%	95%	90%	92%	91%	100%
3	purplewave	57.70%	42%	45%		84%	60%	29%	49%	82%	70%
4	steamhammer	50.11%	3%	8%	16%		50%	62%	67%	97%	99%
5	mcrave	49.81%	14%	5%	40%	50%		89%	65%	35%	99%
6	microwave	43.44%	5%	10%	71%	38%	11%		56%	61%	96%
7	dragon	42.28%	6%	8%	51%	33%	35%	44%		83%	78%
8	ualbertabot	30.45%	1%	9%	18%	3%	65%	39%	17%		91%
9	infestedartosis	8.28%	0%	0%	30%	1%	1%	4%	22%	9%

The same as the official crosstable except for the presentation. Microwave upset PurpleWave even more strongly than last year. And McRave and UAlbertaBot are both carryovers, so of course UAlbertaBot upset McRave like last year.

#	bot	overall	bana	star	purp	stea	mcra	micr	drag	ualb	infe
1	bananabrain	10:47 14:37		17:22 14:46	11:33 12:16	11:02 8:01	11:46 18:46	11:09 5:48	9:02 22:59	10:25 8:52	6:04 -
2	stardust	10:43 16:51	14:46 17:22		14:10 18:06	11:01 14:57	11:57 16:54	10:07 6:10	10:40 24:26	9:36 6:13	9:58 -
3	purplewave	10:26 12:02	12:16 11:33	18:06 14:10		10:14 9:19	11:10 13:18	9:28 7:20	10:06 16:15	10:08 10:47	9:34 8:55
4	steamhammer	7:02 10:45	8:01 11:02	14:57 11:01	9:19 10:14		8:14 8:50	13:42 11:23	12:26 19:05	8:56 11:36	5:09 12:25
5	mcrave	9:15 11:27	18:46 11:46	16:54 11:57	13:18 11:10	8:50 8:14		8:22 6:04	20:43 21:33	12:49 9:09	5:47 9:03
6	microwave	7:03 10:05	5:48 11:09	6:10 10:07	7:20 9:28	11:23 13:42	6:04 8:22		16:35 25:12	8:50 7:06	5:03 10:40
7	dragon	15:45 12:05	22:59 9:02	24:26 10:40	16:15 10:06	19:05 12:26	21:33 20:43	25:12 16:35		12:09 8:39	11:28 12:40
8	ualbertabot	6:54 10:08	8:52 10:25	6:13 9:36	10:47 10:08	11:36 8:56	9:09 12:49	7:06 8:50	8:39 12:09		6:28 8:37
9	infestedartosis	9:52 6:31	- 6:04	- 9:58	8:55 9:34	12:25 5:09	9:03 5:47	10:40 5:03	12:40 11:28	8:37 6:28

The top number in each cell is the median duration of winning games for the bot on the left, and the bottom number is the median duration of losses. Winning fast and losing slowly suggests rush strategies or timing attacks. The opposite suggests cautious play and winning by defense into strong macro, or by gradually outplaying the opponent.

For example, Stardust likes to win with one big attack, and often wins faster than it loses—though it depends on the opponent. UAlbertaBot rushes with every race, and it shows in the numbers. Dragon likes to win by slowly pushing its opponents over. PurpleWave is comparatively balanced.

	overall	vT	vP	vZ	vR
terran	42%		21%	48%	83%
protoss	75%	79%		83%	91%
zerg	38%	52%	17%		50%
random	30%	17%	9%	50%

Um, yeah, protoss came out on top. I think we knew that.

#	bot	overall	vT	vP	vZ	vR
1	bananabrain	84.54%	94%	53%	94%	99%
2	stardust	83.43%	92%	53%	94%	91%
3	purplewave	57.70%	49%	44%	61%	82%
4	steamhammer	50.11%	67%	9%	70%	97%
5	mcrave	49.81%	65%	20%	79%	35%
6	microwave	43.44%	56%	29%	48%	61%
7	dragon	42.28%	-	21%	48%	83%
8	ualbertabot	30.45%	17%	9%	50%	-
9	infestedartosis	8.28%	22%	10%	2%	9%

For me, the lesson of the per-race table is that Steamhammer is very bad against protoss. Also, if it weren’t for the relative performance against UAlbertaBot, McRave would have come out well ahead of Steamhammer.

The AIIDE 2023 tournament has been declared complete. Not all details of the results page are updated, but the basic info is there. In particular, the plain text detailed results file is not there yet, so I can’t run my analysis programs.

#1	BananaBrain	84.54%
#2	Stardust	83.43%
#3	PurpleWave	57.70%
#4	Steamhammer	50.11%
#5	McRave	49.81%
#6	Microwave	43.44%
#7	Dragon	42.28%
#8	UAlbertaBot	30.45%
#9	InfestedArtosis	8.28%

Participation was thin this year, with only 9 entrants. It remains unfortunate that #7 Dragon is the only terran.

Two surprises for me. #1 BananaBrain pulled ahead of #2 Stardust about round 145, and stayed ahead. Stardust does some learning, but after round 50 its performance gradually decreased. After Stardust’s great improvements, I felt pretty sure it would come out ahead. And Stardust did very well—it had a plus score against every opponent. BananaBrain looks impressive, though, with a bigger winning margin against all other opponents except McRave.

#3 PurpleWave became the best of the also-rans, rather than a contender. In the graph, it groups with the middle bots, #3 to #8.

The other surprise: #5 McRave was ahead of #4 Steamhammer for most of the tournament. The main cause was their head-to-head results, not performance against other bots. McRave is a carryover, and I thought my improvements would not have much effect in ZvZ, so I expected results like last year’s, when Steamhammer won head-to-head by a consistent 54%. Steamhammer should even do a little better, because I updated the preparation. As it happened, the head-to-head results varied over the tournament run. Steamhammer was ahead at first, then McRave figured out how to play against it and took the lead, both head-to-head and overall. Only late in the tournament did Steamhammer find an answer and reverse the trend. In the end, the head-to-head score was 49% for Steamhammer, enough to edge it into an overall lead.

Did I introduce a weakness, or was the surprise due to the vagaries of learning? So far, as I prepare the bugfix version, I haven’t found any bugs or weaknesses that seem serious enough to explain it. But data to find the real cause is not released yet.

#6 Microwave and #7 Dragon also had a close race, but in my eyes it doesn’t have the drama of the other two close races.

#8 UAlbertaBot continues to score better than I expected, though from the beginning I thought it was clear what place it would end up in. This year its overall score was 30.45% compared to 43.90% last year, but that’s mainly because there were fewer weak bots for it to beat up on.

Congratulation to #9 InfestedArtosis for participating! You got through with zero crashes, unlike some stronger bots, and scored wins against tough opponents, even if not many wins. Only #1 BananaBrain and #2 Stardust were able to shut you out. The signs say you have a solid foundation to build on.

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.

I was excited to learn that the AIIDE 2023 results are coming in live, this year. Though I see that the detailed results and win percentage over time are not updated as often as the overall results. My first impression is that the ranking so far is exactly what I expected it to be. Looking a little closer, McRave is doing substantially better versus Dragon than last year. (That is last year’s physical hardware version, to match this year.) Both are carryovers, so the difference may be due to the maps, or to the luck of learning. Well, the tournament is not far along. We’ll see how it goes with more data.

It looks like I was wrong about Steamhammer versus BananaBrain. Steamhammer is not as successful there as I expected. Either my tests were unrealistic in some way, or else I introduced a late bug. Or maybe BananaBrain was improved in a way that makes it harder for Steamhammer to do early damage. Steamhammer should put up a good fight in any game where it can kill some probes early on.

I will upload Steamhammer 3.6 to SSCAIT shortly.

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.

The AIIDE 2023 participants have been announced on a Google spreadsheet. They are the usual suspects and the carryovers plus Infested Artosis.

Infested Artosis by Brad Ewing is not high ranked, but I like it. It’s simple. You can read through the code on GitHub in a short time. It’s a far cry from the complex code of Steamhammer, which must be baffling to those who first look at it. At the same time, it has the nice feature that it independently learns what build to play and what unit mix to make. Both are straightforward picks from short lists. There’s provision for learning whether to make static defense too.

Enjoy the competition, Infested Artosis!

In other news, JyJ versus Soulkey in ASL 16 today was a dramatic TvZ best-of-five. The last two games were particularly hard-fought.

Purple Dan Gant did some analysis of the AIIDE 2023 map pool and kindly e-mailed info to me. Here it is; I edited it slightly for clarity.

• Ran games between PurpleWave and BananaBrain across all the AIIDE maps (based on what I think the legal maps are)
• All the maps ran (they were 1.16.1-compatible)
• Some of the maps in my zip appear to be observer maps which are unusable for bot play because they have extra start locations. All but Crossing Field have acceptable alternative versions in the pool
• PurpleWave worked correctly on all maps. Its behavior on Katrina was dumb but explicable due to the backyard natural

He offers results on a google spreadsheet: AIIDE 2023 Map Info (you have to switch between the About and Data sheets). And he created a new version of the map pool which may better represent what we should actually get: Dan’s unofficial copy of AIIDE 2023 map pool version 2.

Thanks, Dan! You have done a public service.

Steamhammer can play on all the maps, but on the ICCup observer maps it feels a need to scout the observer “bases” to make sure nobody’s there. It knows better than to try to expand there. There are maps where it can correctly reject observer slots as non-bases, but it depends on how they are set up. Steamhammer also plays stupidly on Katrina, and I don’t intend to fix it yet because it’s too much time cost to pay down a small risk. At some point I’ll teach it to do the terrain connectivity analysis from first principles, and create a graph that it can reason about. Playing on Outsider is a goal, but one step at a time.

Steamhammer work is coming along nicely. The new feature I’m adding is coming up to the finish line, and it is passing its tests so far. It won’t be fully refined yet, but still a good improvement. I should have time to add one more important feature before the tournament. Between them they will make the bot sharper at taking advantage of opportunities and more resilient when under pressure. Can’t ask more than that in a short timeframe.

As you might guess from the CoG results, Stardust has gained strong new skills since last year. Against zerg in particular, one skill is that it uses corsairs heavily. Another is that it learned the forge expand opening. And yet it can still be defeated.

CherryPi is from 2017, but it shows how. The SSCAIT game Stardust v CherryPi is instructive (I kept the replay for when SSCAIT recycles it).

The recipe is:

1. Facing forge expand, play a greedy opening. CherryPi opened pool first (with a rare 13 pool) so that it could make zerglings if it needed to, but cannons were not going to walk across the map to attack, so it followed up with drones and hatcheries.

2. Fight efficiently in the middle game. CherryPi could not take any one-sided victories against such a tough opponent, but it traded well and cut the protoss army down to a safe size, where zerg could spawn enough defensive units in the time the protoss would take to cross the map.

3. With that breathing room, zerg could safely pull ahead in workers. Then it was just mass and smash.

See how easy it is? It’s a simple matter of being good at everything!

Of course it was only possible because Stardust showed weaknesses. One is that it was cautious and clumsy with its corsairs, and put on less counter-air pressure than it could have. First it kept them with the army, then it flew them over hydras.