Starcraft AI blog

McRave and BananaBrain both recorded all 157 of their mutual games. I chose to put McRave down the left side of the strategy cross, because its longer strategy names make the table hard to read otherwise. I also trimmed off the “PvZ_” and “Z_” prefixes from BananaBrain’s strategy names for compactness.

Blue is good for McRave, red is good for BananaBrain.

mcrave strategies versus bananabrain strategies

	overall	10/12gate	1basespeedzeal	2basespeedzeal	4gate2archon	5gategoon	9/9gate	9/9proxygate	bisu	neobisu	sairdt	sairgoon	sairreaver	stove
overall	28/157 18%	8/54 15%	1/3 33%	1/5 20%	1/1 100%	1/1 100%	1/1 100%	2/8 25%	1/5 20%	1/4 25%	1/3 33%	2/12 17%	1/1 100%	7/59 12%
HatchPool,12Hatch,2HatchMuta	16/86 19%	3/28 11%	1/1 100%	1/3 33%	1/1 100%	1/1 100%	1/1 100%	0/1 0%	-	1/4 25%	0/2 0%	2/10 20%	1/1 100%	4/33 12%
HatchPool,12Hatch,2HatchSpeedling	0/1 0%	-	-	-	-	-	-	0/1 0%	-	-	-	-	-	-
HatchPool,9Pool,2HatchSpeedling	2/2 100%	-	-	-	-	-	-	2/2 100%	-	-	-	-	-	-
PoolHatch,9Pool,2HatchMuta	3/15 20%	0/2 0%	-	-	-	-	-	-	0/3 0%	-	-	-	-	3/10 30%
PoolHatch,9Pool,2HatchSpeedling	0/3 0%	-	-	-	-	-	-	0/3 0%	-	-	-	-	-	-
PoolHatch,9Pool,3HatchMuta	4/19 21%	4/13 31%	-	0/1 0%	-	-	-	-	0/1 0%	-	-	-	-	0/4 0%
PoolHatch,9Pool,6HatchHydra	0/2 0%	-	-	-	-	-	-	-	-	-	-	0/1 0%	-	0/1 0%
PoolHatch,Overpool,2HatchMuta	0/4 0%	0/2 0%	0/1 0%	-	-	-	-	-	-	-	-	0/1 0%	-	-
PoolHatch,Overpool,2HatchSpeedling	0/1 0%	-	-	-	-	-	-	0/1 0%	-	-	-	-	-	-
PoolHatch,Overpool,3HatchMuta	3/14 21%	1/9 11%	0/1 0%	-	-	-	-	-	1/1 100%	-	1/1 100%	-	-	0/2 0%
PoolHatch,Overpool,6HatchHydra	0/10 0%	-	-	0/1 0%	-	-	-	-	-	-	-	-	-	0/9 0%

There we have the explanation for the 2 lonely HatchPool,9Pool,2HatchSpeedling games: The strategy was a successful reaction to proxy gates. I read it as meaning that the build is 9 hatch, 9 pool, and gas soon for the zergling speed upgrade.

BananaBrain split its effort between 10-12 gate and the Stove (a scout into dark templar build), very different builds. McRave answered both mostly with 12 hatch into 2 hatch muta. A hydralisk opening would have been a more natural way to counter both, but play what you’re good at.

mcrave as seen by bananabrain

mcrave played	#	bananabrain recognized
HatchPool,12Hatch,2HatchMuta	86	84 12hatch | 2 unknown
HatchPool,12Hatch,2HatchSpeedling	1	1 unknown
HatchPool,9Pool,2HatchSpeedling	2	2 12pool
PoolHatch,9Pool,2HatchMuta	15	8 9pool | 7 overpool
PoolHatch,9Pool,2HatchSpeedling	3	2 overpool | 1 9pool
PoolHatch,9Pool,3HatchMuta	19	13 9pool | 6 overpool
PoolHatch,9Pool,6HatchHydra	2	1 overpool | 1 9pool
PoolHatch,Overpool,2HatchMuta	4	4 overpool
PoolHatch,Overpool,2HatchSpeedling	1	1 overpool
PoolHatch,Overpool,3HatchMuta	14	14 overpool
PoolHatch,Overpool,6HatchHydra	10	10 overpool

BananaBrain was accurate in recognizing 12 hatch and overpool, but had trouble with 9 pool. It did not try to narrow the build down any further than that.

bananabrain as seen by mcrave

bananabrain played	#	mcrave recognized
10/12gate	54	31 2Gate,10/12,Corsair | 8 2Gate,10/12,DT | 4 2Gate,10/12,ZealotRush | 3 2Gate,10/17,Corsair | 2 2Gate,Unknown,Corsair | 2 2Gate,9/9,DT | 2 2Gate,9/9,Corsair | 1 2Gate,10/17,4Gate | 1 2Gate,10/17,DT
1basespeedzeal	3	2 1GateCore,2Zealot,DT | 1 1GateCore,2Zealot,Corsair
2basespeedzeal	5	3 FFE,Forge,Speedlot | 1 FFE,Nexus,Speedlot | 1 FFE,Nexus,5GateGoon
4gate2archon	1	1 FFE,Forge,5GateGoon
5gategoon	1	1 FFE,Nexus,5GateGoon
9/9gate	1	1 2Gate,9/9,Corsair
9/9proxygate	8	7 2Gate,Proxy,ZealotRush | 1 2Gate,9/9,Unknown
bisu	5	2 FFE,Forge,Unknown | 2 FFE,Forge,5GateGoon | 1 FFE,Nexus,Unknown
neobisu	4	4 FFE,Forge,Speedlot
sairdt	3	3 1GateCore,2Zealot,Corsair
sairgoon	12	6 FFE,Forge,5GateGoon | 2 FFE,Nexus,5GateGoon | 2 FFE,Forge,Unknown | 1 FFE,Gateway,5GateGoon | 1 FFE,Gateway,Unknown
sairreaver	1	1 FFE,Forge,Unknown
stove	59	27 1GateCore,2Zealot,Corsair | 10 1GateCore,Unknown,Corsair | 6 1GateCore,Unknown,DT | 6 1GateCore,2Zealot,DT | 3 2Gate,10/12,DT | 3 2Gate,10/12,ZealotRush | 2 2Gate,10/17,4Gate | 1 2Gate,10/17,DT | 1 2Gate,10/12,4Gate

In 1-base protoss plays, McRave tried to distinguish when the gates were made, and often got it right but had some trouble. It seems like something you can’t do perfectly, even if you combine direct scouting of the gates with inferences based on the enemy army. Recognizing the enemy build precisely doesn’t seem possible in general, though you can usually get close.

I’m taking the bots in finishing order; McRave is next. Last year I analyzed McRave’s three-part strategy representation and learning algorithm. These apparently have not changed in outline, though details may have changed. The set of available strategies has been updated. For example, 6HatchHydra is new this year. It follows that the set of enabled strategy triples has also changed.

McRave is much stronger this year. It has become noted for dangerous mutalisk control.

9 pool into 2 hatch muta worked best, with 3 wins out of 22. That is not intuitive. The more natural 12 hatch into 2 hatch muta was tried more but was less successful. Did Stardust react inefficiently to the 9 pool? McRave appears to correctly understand that Stardust ends up with 4 gates despite taking different routes to get there. That’s kind of impressive.

Last year McRave scored 3 out of 150 against Stardust. This year it scored 8 out of 150 against a stronger Stardust—which seems to have updates specifically to defeat McRave, since McRave was the only bot to upset it in CoG 2021. Good progress!

McRave’s wins over BananaBrain are dominated by games where BananaBrain timed out. In game 2754, BananaBrained timed out with 320 frames over 55ms, when McRave had 318 frames over 55ms—close call! McRave had more timeouts than any other bot, but only 3 losses to BananaBrain by timeout. BananaBrain’s timeouts seem to be concentrated on Dragon and McRave, and to a lesser extent on Steamhammer.

If HatchPool,9Pool,2HatchSpeedling was tried twice and won twice, why wasn’t it tried more often? The first try was on game 30 of 0-156. I imagine that it was a reactive build, not enabled as an initial choice but switched to under given circumstances. I didn’t read the source to verify that.

The enemy table shows a complex set of strategies by BananaBrain.

Over half of McRave’s losses to Dragon were by timeout. I think Dragon is an especially easy bot to time out against, because its strong macro and big battles with light units put heavy demands on the opponent.

The enemy table shows 89 games with Unknown,Unknown,Unknown. Apparently Dragon often denied scouting. Presumably the scouting overlord was afraid to approach due to marines, and any scouting drone was turned away. Also, I wonder about 2Rax,Proxy,Unknown. Did Dragon really proxy once, or was it a misrecognition? On Python, bases can be close by air. If McRave measures proxy distance by air distance, it might take a barracks in the enemy main for a proxy.

McRave chose from the same fixed set of 3 strategies against all the zergs. Only the 1 hatch mutalisks were able to hold their own with Steamhammer.

Hurrying the mutas too much did not help against WillyT. The enemy table shows that WillyT sometimes countered with goliaths. Does McRave later make a hydra switch to fight the goliaths? WillyT sometimes goes for goliaths with 2 tanks, and it’s sensible to fight back with hydralisks. I didn’t see a hydra switch in the games I watched.

I noticed that McRave doesn’t clear the terran scout from its main until mutas come out. WillyT gets to know the exact timing for its turrets with no need to spend a scan.

The mutalisks did in DaQin. DaQin’s slow start puts 12 hatch ahead of other choices; DaQin makes cannons before nexus regardless of what the opponent does. DaQin defends its natural entrance with cannons, but not its nexus, so the mutalisks have a free hand and DaQin finds itself short of probes.

McRave met UAlbertaBot with the same strategies as last year (except for one stray PoolHatch,12Pool,2HatchMuta this year). The same strategies by name, that is. The actual play was different and performed far worse against UAlbertaBot’s pressure builds. I looked at some games. When McRave respected its enemy and defended itself, it generally won. Sometimes it seemed to arrogantly conclude “Pff, you’re not worth spending a sunken on” and got overrun. As far as I could tell from watching games, it wasn’t a scouting miss—though it’s easy to overlook things in watching games. It had the feel of a bug.

The submitted Steamhammer was mistakenly configured to retain 100 game records per opponent. I had thought it was set for 200, and didn’t double-check. So of the 157 games against each opponent, of which 150 counted in the tournament, I have records for only the final 100. That’s 93 tournament games plus the 7 extra at the end.

My prepared data was successful. For all opponents which I prepared for, the prepared openings were among the highest scoring (including the zero score versus Stardust). It’s notable that Steamhammer’s gas steal was not successful against any opponent, perhaps another sign of an elite tournament. It was either infeasible or abandoned as a failure against every opponent except DaQin, and did no good then.

Steamhammer’s game records are rich with data. To show a little bit more of it, I added a new feature in the opening table. There are new “wins” and “losses” columns showing the median time that winning and losing games with that opening lasted. The median is a better measure than the mean, because we can expect the distribution of game times to be right-tailed: Games are limited to between zero minutes and one hour, but we expect a hump nearer to zero and a long tail of slower games. That inflates the mean and makes it misleading. For the tournament, I turned off surrendering, so Steamhammer played its losses out to the end.

Steamhammer lost every game, but there is still valuable info here. If you’re losing all games, the game duration is a plausible proxy for how much trouble you caused the opponent. Especially so if you tried a rush opening and ended up in a long game—either the rush did some damage, or the opponent overreacted and was slowed down. Here, a couple of 7 pool builds were among the longest games. Steamhammer probably should have repeated them.

Notice the 4 pool and the hive rush. Steamhammer tried the whole range. Steamhammer recognized Stardust’s build in 2 games as nexus without cannons, a reaction that Stardust did not have last year. Otherwise, results are similar to last year’s.

In the 150 tournament games, Steamhammer scored 25 wins versus BananaBrain. Of those, 15 were due to BananaBrain suffering a frame timeout. Ouch. The game scores say that BananaBrain was ahead in 11 of the 15 games when it timed out. So the win percentages and times need to be interpreted carefully. The wins overall were longer games than the losses, possibly because BananaBrain was more likely to time out in a longer game with a larger game state to model and more units to control.

11HatchTurtleLurker scored over 50% in 15 games! Is it particularly good at prompting BananaBrain to time out? If I’d known about it ahead of time, I could have added it to my preparation and perhaps scored higher.

The build 2x10HatchSlow is shown as tried 7 times with no wins. I know from watching games that the opening scored wins earlier in the tournament, before the final 100 games; that is why it was tried so often later on. The build is very similar to Broken Horn’s 10 hatch-9 hatch-pool, but (I think) slightly more efficient. Apparently BananaBrain learned to avoid lines that lose to the mass of slow zerglings.

Successfully stealing its gas caused BananaBrain to take its gas sooner. I haven’t seen that before. In any case, it was only 3 games; Steamhammer found the gas steal unprofitable.

The most successful openings were 5HatchPool (5 hatcheries before pool, a supremely greedy build to exploit bots that never attack early) and ZvT_3HatchMutaExpo, the two openings I selected as preparation. For bots carried over from the previous year, good preparation is easier.

Dragon has a chaotic play style. Steamhammer’s wildest game of the tournament may be Steamhammer-Dragon on Longinus (replay file). Dragon played a V strategy: Vultures, valkyries, and vessels.

Again two of my prepared builds, 9PoolHatchSpeedAllIn and ZvZ_OverpoolTurtle, were the top choices. Both are tough for most zerg bots to handle. My other prepared build, ZvZ_Overgas9Pool, does not appear in these 100 games. Apparently it flopped and was abandoned early. Rushy builds ended up winning faster than they lost, and more macro builds were the reverse, as you might expect. The timing table shows that McRave went spire nearly every game (overlords do not count as “air units” there), and not slowly. That’s normal for ZvZ, of course, but it shows that McRave did not favor builds to overrun the opponent with zerglings.

WillyT has become much stronger over the past year. It is better at handling Steamhammer’s lurker builds—except for the especially early 9 pool lurker build, which apparently catches it unready. Steamhammer’s improvements in lurker play were important to keep up with progress. I think Steamhammer’s diverse mix of openings was essential to counter WillyT, which has its own diverse mix and will figure out how to counter anything that is too predictable.

Steamhammer’s closest game of the tournament was Steamhammer-WillyT on Empire of the Sun (replay file). Steamhammer decisively stopped WillyT from taking the nearby north island base, but allowed it to hold the distant south island despite scouting it the instant it started. Notice WillyT’s interesting but somewhat uncoordinated dropship play throughout the game.

Microwave had too many weaknesses. Of the 4 openings with 80% plus win rates, 3 were from preparation and one was Steamhammer’s discovery during the tournament. It’s interesting that the 12 hatch build ZvZ_12HatchMain was faster to win than to lose. I think that means it won with zerglings from its extra larvas.

The plan table shows that Microwave followed its own broad range of plans. In the timing table, see the wide and matching variation in Microwave’s gas timing and air unit timing. Did Microwave never get zergling speed in long games?

After this upset, I think I’ll take DaQin as a test opponent and finally figure out the skills to defeat it. DaQin is a Locutus fork, so beating it probably means doing better against other protoss bots.

The timing table shows that DaQin was remarkably late with air units. That includes both corsairs and observers—DaQin was late with both of them. In fact, I don’t remember whether it makes corsairs at all. Mutalisks might be a good choice to win.

When I was preparing opponent-specific data, Steamhammer had an overwhelming score against FreshMeat on BASIL. This result is good but not overwhelming; FreshMeat improved a lot in a short time. I had recognized that FreshMeat had made great strides, but there was not enough recent data to show what was working in the most recent games. So I made no preparation at all. These tables show an example of how Steamhammer figures out an opponent from scratch. I think it did OK.

Comparing this year to last year, Steamhammer actually did a little worse against UAlbertaBot. The skills I improved over the last year didn’t include skills to defeat UAlbertaBot’s pressure builds, or to adapt better to its random race.

BananaBrain and Dragon both recorded their own opening builds for all 157 games played, so I can align their learning files and see how their strategies matched up against each other. BananaBrain also recorded its representation of what the opponent played, so I can compare its idea of Dragon’s build with Dragon’s own idea. I first did this last year. Dragon is carried over from last year unchanged, while BananaBrain is much stronger now.

The win rates and coloring are from the point of view of BananaBrain. Blue is good for BananaBrain and red is good for Dragon.

bananabrain strategies versus dragon strategies

	overall	1rax fe	2rax bio	2rax mech	bio	dirty worker rush	mass vulture	siege expand
overall	117/157 75%	8/8 100%	19/30 63%	8/8 100%	12/13 92%	8/8 100%	27/36 75%	35/54 65%
PvT_10/12gate	34/48 71%	1/1 100%	9/17 53%	1/1 100%	2/2 100%	2/2 100%	3/5 60%	16/20 80%
PvT_1gatedtexpo	0/1 0%	-	0/1 0%	-	-	-	-	-
PvT_28nexus	3/6 50%	-	1/2 50%	-	-	-	1/1 100%	1/3 33%
PvT_2gaterngexpo	2/4 50%	-	0/1 0%	-	-	-	1/1 100%	1/2 50%
PvT_32nexus	0/1 0%	-	-	-	-	-	-	0/1 0%
PvT_9/9gate	78/96 81%	7/7 100%	9/9 100%	7/7 100%	10/11 91%	6/6 100%	22/29 76%	17/27 63%
PvT_9/9proxygate	0/1 0%	-	-	-	-	-	-	0/1 0%

dragon as seen by bananabrain

dragon played	#	bananabrain recognized
1rax fe	8	7 T_unknown | 1 T_fastexpand
2rax bio	30	30 T_unknown
2rax mech	8	8 T_unknown
bio	13	13 T_unknown
dirty worker rush	8	8 T_unknown
mass vulture	36	21 T_1fac | 14 T_unknown | 1 T_2fac
siege expand	54	38 T_1fac | 16 T_unknown

Last year this table showed that BananaBrain was weak at recognizing Dragon’s builds, with a lot of unknowns. There are more recognized builds this year, but BananaBrain plays differently so I’m not sure whether BananaBrain has improved at recognition. What is clear is that everything is blue. Recognizing some builds does not seem to have helped BananaBrain; it did well no matter what.

Dragon records for each game only its own build and win/loss, so the information is sparse. It has a total of 7 builds. Dragon is a carryover bot, and I analyzed its game records from AIIDE 2020 last year. Dragon considers that “the opening” is a very brief phase of the game: It quickly adapts to what it sees of the opponent’s play, and the opening build fades out of view. Last year I found that, against many opponents, the choice of opening build made little difference; the game was decided later.

It’s interesting that the only openings to make a dent were “2rax bio” and “bio”. Was Stardust surprised by marines? If Stardust made zealots to get units faster, that may have backfired.

Again, the marines were a relatively successful choice against protoss. It’s a surprise. #8 DaQin below is different. Yesterday we saw that BananaBrain liked zealot openings against Dragon, and it’s true that marines with good micro can hold their own against zealots. Maybe BananaBrain liked zealots because they upset Dragon’s tech builds, and Dragon found that marines answered best, so that the two settled into this equilibrium with neither bot able to 100% counter the other. It’s a nice story, at least.

Switching between opposite builds like fast expand (“1rax fe”) and worker rush (“dirty worker rush”) is not a bad plan for defeating Steamhammer. I don’t think Dragon did it on purpose, though. Most openings scored about equal.

Again, most were about equal, with only a couple of exceptions. “1rax fe” was also best against last year’s McRave, even though it played rather differently.

Last year Dragon chose “2rax mech” as its build to trample on WillyT (winning fewer, 94%, even though this year’s WillyT is substantially stronger). I think it found something that worked and felt no need to experiment any further.

Best was the fast expansion. (“1rax fe” is faster than “siege expand”.) That makes sense against DaQin’s style of play.

Mostly about equal again. At some point I’ll look at the games and see how FreshMeat upset Dragon.

Here’s my summary of BananaBrain’s learning files. BananaBrain records both its own strategy and the recognized enemy strategy for every game.

#1 stardust

opening	games	wins	first	last
PvP_10/12gate	5	0%	9	121
PvP_12nexus	5	0%	10	119
PvP_2gatedt	5	0%	3	120
PvP_2gatedtexpo	16	19%	0	125
PvP_2gatereaver	5	0%	1	118
PvP_3gaterobo	5	0%	13	123
PvP_3gatespeedzeal	5	0%	5	116
PvP_4gategoon	24	25%	12	126
PvP_9/9gate	5	0%	6	122
PvP_9/9proxygate	38	29%	8	156
PvP_nzcore	13	15%	4	149
PvP_zcore	5	0%	7	117
PvP_zcorez	9	11%	11	144
PvP_zzcore	17	24%	2	127
14 openings	157	17%

enemy	games	wins
P_1gatecore	6	50%
P_2gate	26	19%
P_2gatefast	13	31%
P_4gategoon	107	14%
P_cannonturtle	1	0%
P_unknown	4	0%
6 openings	157	17%

It occurred to me a little late that many people would find the Stardust data table easier to understand if the frame times were converted to minutes and seconds. So here’s that version. See the previous post from today.

		firstDarkTemplarCompleted				pylonInOurMain				firstMutaliskCompleted
opponent	games	n	min	median	max	n	min	median	max	n	min	median	max
bananabrain	155	20	5:15	5:29	16:11	0	-	-	-	0	-	-	-
dragon	156	0	-	-	-	0	-	-	-	0	-	-	-
steamhammer	158	0	-	-	-	0	-	-	-	17	4:59	5:43	7:11
mcrave	157	0	-	-	-	0	-	-	-	124	6:17	7:35	11:12
willyt	157	0	-	-	-	0	-	-	-	0	-	-	-
microwave	157	0	-	-	-	0	-	-	-	17	5:07	5:55	7:54
daqin	156	126	5:13	5:29	12:36	2	1:53	1:54	1:55	0	-	-	-
freshmeat	157	0	-	-	-	0	-	-	-	1	11:40	11:40	11:40
ualbertabot	157	17	4:19	4:29	4:36	0	-	-	-	0	-	-	-

I investigated how Stardust’s learning works, and what it learned. It’s unusual, so it was worth a close look.

In its learning file of game records for each opponent, Stardust records values for 3 keys for each game, firstDarkTemplarCompleted, pylonInOurMain, and firstMutaliskCompleted. If the event occurs in the game, the value is the frame time of the event; otherwise the value is 2147483647 (INT_MAX, the largest int value, in this C++ implementation). It also records whether the game was a win or a loss. It records the hash of the map, too, but that doesn’t seem to be used again.

summarizing the data

The class Opponent is responsible for providing the learned information to the rest of the bot. It summarizes the game records via two routines.

  int minValueInPreviousGames(const std::string &key, int defaultNoData, int maxCount = INT_MAX, int minCount = 0);

If there are at least minCount games, then look through the game records, most recent first, for up to maxCount games. Look up the key for each game and return its minimum value, or the default value if there are none. This amounts to finding the earliest frame at which the event happened, or the default if it did not happen in the specified number of games.

   double winLossRatio(double defaultValue, int maxCount = INT_MAX);

Look through the game records, most recent first, for up to maxCount games and return the winning ratio, or the default value if there are no games yet.

using the summarized data

Each of the 3 keys is used in exactly one place in the code. Here is where firstDarkTemplarCompleted is looked up in the PvP strategy code:

    if (Opponent::winLossRatio(0.0, 200) < 0.99)
    {
        expectedCompletionFrame = Opponent::minValueInPreviousGames("firstDarkTemplarCompleted", 7300, 15, 10);
    }

This means “If we’re rolling you absolutely flat (at least 99% wins in the last 200 games), then it doesn’t matter. Otherwise there’s some risk. In the most recent 15 games, find the earliest frame that the first enemy dark templar was (estimated to be) completed, or return frame 7300 if none.” The default frame 7300 is not the earliest a DT can emerge; they can be on the map over a thousand frames earlier. So it is not a worst-case assumption. Further code overrides the frame number if there is scouting information related to dark templar production. It attempts to build a defensive photon cannon just in time for the enemy DT’s arrival, and sometimes to get an observer.

The key pylonInOurMain is part of cannon rush defense. Stardust again checks the win ratio and again looks back 15 games with a minimum game count of 10, this time with a default of 0 if there are not enough games. It starts scouting its base 500 frames (about 21 seconds) ahead of the earliest seen enemy pylon appearing in its base, which may be never. The idea is that Stardust doesn’t waste time scouting its own base if it hasn’t seen you proxy a pylon in the last 15 games, and delays the scout if the pylon is proxied late.

The key firstMutaliskCompleted is used very similarly, to decide whether and when to defend each nexus with cannons. The goal is to get cannons in time in case mutalisks arrive without being scouted. There are simple rules to decide how many cannons at each nexus:

    // Main and natural are special cases, we only get cannons there to defend against air threats
    if (base == Map::getMyMain() || base == Map::getMyNatural())
    {
        if (enemyAirUnits > 6) return 4;
        if (enemyAirThreat) return 3;
        if (enemyDropThreat && BWAPI::Broodwar->getFrameCount() > 8000) return 1;
        return 0;
    }

    // At expansions we get cannons if the enemy is not contained or has an air threat
    if (!Strategist::isEnemyContained() || enemyAirUnits > 0) return 2;
    if (enemyAirThreat || enemyDropThreat) return 1;
    return 0;

If the firstMutaliskCompleted check says that it’s time, it sets enemyAirThreat to true and makes 3 cannons each at main and natural, and at least 1 at each other base.

the data itself

Here’s my summary of the data in Stardust’s files. The files include prepared data. I left the prepared data out; this covers only what was recorded during the tournament. The tournament was run for 157 rounds, although the official results are given after round 150. The table here is data for all 157 rounds. I don’t have a way to tell which unrecorded games were from rounds 1-150 and which were from 151-157... though I think I could guess.

n is the number of games for which a value (other than 2147483647) was recorded for the key. The values are frame numbers.

		firstDarkTemplarCompleted				pylonInOurMain				firstMutaliskCompleted
opponent	games	n	min	median	max	n	min	median	max	n	min	median	max
bananabrain	155	20	7579	7897.5	23319	0	-	-	-	0	-	-	-
dragon	156	0	-	-	-	0	-	-	-	0	-	-	-
steamhammer	158	0	-	-	-	0	-	-	-	17	7188	8241	10355
mcrave	157	0	-	-	-	0	-	-	-	124	9070	10939	16146
willyt	157	0	-	-	-	0	-	-	-	0	-	-	-
microwave	157	0	-	-	-	0	-	-	-	17	7371	8534	11397
daqin	156	126	7533	7912.5	18154	2	2721	2743.5	2766	0	-	-	-
freshmeat	157	0	-	-	-	0	-	-	-	1	16801	16801	16801
ualbertabot	157	17	6230	6477	6627	0	-	-	-	0	-	-	-

As you might expect after deep contemplation of the nature of reality, only protoss makes dark templar or proxy pylons, and only zerg makes mutalisks. Nothing interesting was recorded for the terran opponents.

Notice that UAlbertaBot sometimes makes dark templar much earlier than the no-data 7300 frame default time; the others do not. DaQin is recorded as twice placing a proxy pylon in Stardust’s main. I didn’t think it ever did that. I guess it’s a holdover from the Locutus proxy pylon play, to trick opponents into overreacting? DaQin made DTs in most games, and McRave went mutalisks in most games. FreshMeat is recorded as having made a mutalisk (or more than one) in exactly one game, which seems unusual.

Before I dig into what each bot learned, I thought I’d look at the win percentage over time graph. Every bot wrote data, and it is likely that every bot attempted to learn and improve over time. Only some succeeded in improving their results, though.

Every bot shows a startup transient on the graph. The early swings up and down are controlled by some combination of luck and learning; luck because there are few games so statistical variation is high, and learning if and when the learning algorithms make fast adjustments (I think they usually do). To disentangle luck from learning, I think I want both statistical tests and a look into the algorithms to see what the learning rates could be. It would be too much for one post. In this post, I’m looking at the curves after 20 or 30 rounds, when the swings have mostly leveled off. I’m answering the question: Is the bot able to keep learning throughout a long tournament, outlearning its competition in the long run?

Four bots more or less held even. There are wobbles or slight trends, but not large ones. It’s what you expect if most bots are about equally good at lifetime learning. The learning systems are more or less saturated, and when one discovers an exploit, its counterpart figures out soon enough how to neuter the exploit, or so I imagine it. The learning competition is near an equilibrium.

Stardust doesn’t learn much, and apparently doesn’t have to. Steamhammer and McRave have messy early curves, perhaps reflecting complicated learning systems. FreshMeat has a beautiful clean early curve, unlike any other bot’s, suggesting that it knows what it is doing and straightforwardly does it. All 3 of the lower bots show low humps followed by slight regressions. I provisionally interpret that as the bot’s learning system saturating, then its opponents adjusting to that over time.

Four bots were able to improve. BananaBrain was in a class by itself, improving far more than any other bot. WillyT, Microwave, and UAlbertaBot had slight upward trends. None of them looks as impressive as AIUR did in 2015.

What gives BananaBrain a steeper curve? Is it good at learning in the long term, or bad at learning in the short term? (See that down-hook at the beginning.) I’ll look into it later on.

Dragon and DaQin fell behind. If somebody’s going up, somebody else must be going down. It may not be a coincidence that both are carryover bots from last year. Dragon’s learning files have a simple structure, the strategy name and win/loss. DaQin plays few strategies and has few ways to escape from exploits that other bots may find.

Next: Looking at Stardust’s learning.

I looked in each bot’s final write directory to see what files it wrote, if any, and in its AI directory to see if it had prepared data for any opponents. Be sure to note: A bot does not necessarily use the data it writes. Preparation for specific opponents is not necessarily in the form of data in the AI directory, it might be in code.

#	bot	info
1	Stardust	Unlike last year, this year Stardust wrote data. It’s in JSON format, and records the map by hash, win or loss, and the timings of up to 3 game events, named firstDarkTemplarCompleted, firstMutaliskCompleted, and pylonInOurMain. The times look like frame numbers, and the great majority are 2147483647 (-1 printed as unsigned), which must mean “didn’t happen”. There is prepared data for 7 opponents (including PurpleWave which did not compete), so I assume that Stardust uses the data. I’ll find out for sure when I look at the source.
2	BananaBrain	The learning files look unchanged from last year and the year before: One file for each opponent in the form of brief records of results. Each record consists of date+time, map, BananaBrain’s strategy (“PvZ_9/9proxygate”), the opponent’s recognized strategy (“Z_9pool”), a floating point number which we were told last year is the game duration in minutes, and the game result. Pre-learned data for DaQin and Dragon, the two stronger carryover bots. Last year there was pre-learned data for more opponents; maybe prep for opponents that might change turned out risky.
3	Dragon	Simple game records, one per line, with strategy and game result, like "siege expand" won.
4	Steamhammer	Steamhammer’s learning file format is documented here.
5	McRave	The files look to have the same information as last year, but the format is slightly different. Two files for each opponent, named like ZvU UAlbertaBot.txt and ZvU UAlbertaBot Info.txt. The first file is short and counts wins and losses overall and for each of McRave’s strategies. The info file has detailed game records with aspects of the opponent’s strategy (2Gate,Main,ZealotRush), McRave’s strategy at 3 levels of abstraction (PoolHatch,Overpool,2HatchMuta), timings, and unit counts. No prepared files.
6	WillyT	The files seem to have been corrected since last year. There is one file per opponent, one line per game, with lines that look like 20211005,Z,03,0. The items look like date, opponent race, a number 01 02 or 03, and win/loss. No prepared files.
7	Microwave	Result and history files for each opponent. They look identical to last year’s, except that Microwave now lists a much larger number of strategies for itself. The result files count wins and losses for each Microwave strategy. The history files have a one-line record of data about each game. Also pre-learned history files for all opponents, each with over 100 game records.
8	DaQin	Carried over from last year. Learning files straight from its parent Locutus (very similar to the old format Steamhammer files). No prepared files (and they’d be out of date if they existed).
9	FreshMeat	Three files for each opponent, except 6 files for UAlbertaBot, presumably because it plays random. The contents of the files are opaque: Two are bare lists of numbers, one is a list of incomprehensible 14-character strings. I’ll have to read the code. No prepared files.
10	UAlbertaBot	Carried over from past years. For each opponent, a file listing strategies with win and loss counts for each.

The only real surprise is Stardust’s minimalist and rather weird-seeming data. FreshMeat is new, of course, so anything it did would be unsurprising! It’s notable that every single participant wrote learning data, but that’s not a surprise either because this was an elite tournament. Except for Stardust, all the elite bots have used learning for years.

In unrelated news, I expected that CoG would post replays and learning files shortly after the AIIDE submission deadline. But no, they haven’t done it yet.