Starcraft AI blog

I meant to summarize McRave’s learning data today, but to know what to put in the tables I had to understand how the numbers are used. Yesterday I examined McRave’s strategy representation with three elements, like “PoolHatch,Overpool,2HatchMuta”. In the code, the elements are named “build” (like PoolHatch), “opener” (like Overpool) and “transition” (like 2HatchMuta). Today I read the code to see what the numbers in the learning files are and how they are used.

Here’s a sample data file, showing McRave doing well versus Steamhammer. The first two numbers are the overall wins and losses. After that, delimited by dashes, is a section for the first build, followed by a section for the openers of that build and a section for the transitions of the build. Then more sections for the other two builds and their appendages. Each element has an independent count of wins and losses.

86 64
-
HatchPool 0 0
-
12Hatch 0 0
-
2HatchMuta 0 0
2HatchSpeedling 0 0
-
PoolHatch 28 27
-
4Pool 0 0
9Pool 0 0
Overpool 0 0
12Pool 28 27
-
2HatchMuta 16 17
2HatchSpeedling 12 10
3HatchSpeedling 0 0
-
PoolLair 58 37
-
9Pool 58 37
-
1HatchMuta 58 37

The code calls a function to check which triples are allowed and deals with other minor details, but even with the fiddly bits it’s simple: It picks the build with the highest UCB value, then given that build the corresponding opener with the highest UCB value, then given that build and opener the transition with the highest UCB value. Because of how the data file is organized, this can be done in one pass. The code is in the file LearningManager.cpp in the nested function parseLearningFile().

In theory, this three-level hierarchy could speed up learning. For example, you might be able to conclude that PoolHatch is better than PoolLair against some opponent, even if you don’t have enough data to know which PoolHatch opener or transition is best. My intuition is that the hierarchical scheme should on average work better than a flat scheme, but that there will be perverse situations where it does worse. Many of the triples are not allowed, which limits the value of the hierarchy. There should be enough data from this tournament to judge whether the hierarchy brought an advantage; it would be interesting to do the analysis.

Next: OK, now I know what tables to generate. I have to add some features to my script, but soon I should be able to post the summary tables.

McRave lists its strategy choices with three names, such as “PoolLair,9Pool,1HatchMuta.” In trying to understand what McRave learned, I myself learned more about what the 3 names mean: They are chosen from a small structured set of composable elements which are strung together into full strategies. Here is the scheme:

1. HatchPool
1.1.1 12Hatch
1.2.1 2HatchMuta
1.2.2 2HatchSpeedling
2. PoolHatch
2.1.1 4Pool
2.1.2 9Pool
2.1.3 Overpool
2.1.4 12Pool
2.2.1 2HatchMuta
2.2.2 2HatchSpeedling
2.2.3 3HatchSpeedling
3. PoolLair
3.1.1 9Pool
3.2.1 1HatchMuta

I made up the numbering to depict how the elements fit together, as I understand it: First you pick a top-level category of builds, 1 2 or 3. Say you pick 2, spawning pool then hatchery. Under that, you pick a concrete build order that implements the top-level category, 2.1.x, from the list of ways that McRave knows to make a spawning pool first. After that—not under it—you pick a followup strategy, 2.2.x, from the strategies that are reachable via that category of builds. Notice that 9Pool appears under both PoolHatch and PoolLair (both are pool first), and that 2HatchMuta appears under both HatchPool and PoolHatch (all you need are 2 hatcheries). I count 15 different possible strategies in this scheme, 2 under HatchPool, 12 under PoolHatch, and 1 under PoolLair.

In principle, combinatorial schemes like this can represent a huge number of strategies with a small number of building blocks. Adding the one new item of 2.2.4 2HatchLurker would add 4 new strategies, though they might not all make sense in games. Having separate elements could also give a learning algorithm clues about how to generalize its findings.

In practice, McRave does not play all the strategies it can represent. It appears to have tight restrictions based on matchup. This list is based on the strategies that are mentioned in the learning files, not on code analysis, so it might be incomplete or misleading.

• ZvT: HatchPool-12Hatch or PoolHatch-Overpool or PoolHatch-12Pool, then for any of those, 2HatchMuta or 2HatchSpeedling.
• ZvP: PoolHatch-Overpool or PoolHatch-12Pool, then 2HatchMuta or (rarely) 2HatchSpeedling.
• ZvZ: PoolHatch-12Pool then 2HatchMuta or 2HatchSpeedling, or PoolLair-9Pool-1HatchMuta.
• ZvR: PoolHatch-Overpool or PoolHatch-12Pool, then 2HatchMuta or 2HatchSpeedling.

Next: Tables showing what strategy choices McRave learned.

Dragon’s learning file format is spare, one line for each game giving strategy name and win or loss, nothing more. Dragon has 7 strategies, and against most opponents tried all of them. Its habit is to keep with a winning strategy, trying others sporadically but generally switching when the current plan starts to fail.

Dragon calls its worker rush “dirty worker rush”. Perhaps we should get it together with Stone so it can learn a nice clean worker rush.

As you can see in the “first” column (the first game each strategy was played), Dragon tried all 7 strategies in the first 7 games because they all lost on their first tries. Worker rush turned out to be the most successful plan, as far as that goes, which is very interesting. Mass vultures were the most-played plan despite not having the highest win rate, apparently because the worker rush had a string of losses so that vultures looked better in recent games. (Maybe Dragon figured that Stardust had learned how to deal with the worker rush.)

How did mass vultures have any chance against Stardust’s dragoons? I located a couple of the “mass vulture” wins and watched them. In fact, tanks were the core of Dragon’s army and the vultures acted as buffer. It looked like regular tank-vulture unit mix with regular tank pushes.

Against PurpleWave, and BananaBrain below, most Dragon strategies worked about equally well. Apparently it has well-balanced play against protoss. Actually I think the explanation may be different: Once the opening is over, Dragon quickly adapts to the enemy, playing against the units it sees. If I guess right, then its goal in the opening is to survive in a good position, and after that Dragon will produce whatever units it needs, so the opening doesn’t much affect the outcome. Obviously the worker rush doesn’t leave much room for adaptation, so it is an exception.

This hypothesis explains why Dragon can do well though it records so little data about each game: The openings often don’t much matter.

Fast expand works versus McRave...

2 barracks is good against Microwave...

... but against Steamhammer, again, most strategies look about the same. Watching games, I think Dragon converges on a diverse unit mix fairly quickly after the opening.

I checked out a “mass vulture” game against Steamhammer, and it looked different from the same strategy against Stardust. Dragon made a modest number of vultures and researched spider mines, but added tanks and wraiths and soon the unit mix looked like most Dragon-Steamhammer games.

Most curious: Against ZZZKBot, factory openings predominate. Checking the game durations, most games that ZZZKBot won were short, meaning that it played its 4 pool with success. Most games that Dragon won were longer, so either ZZZKBot did not 4 pool or else Dragon was slow to counterattack after surviving.

Microwave’s history files include both pre-training games and tournament games. I removed the pre-training games, and these tables show only tournament results. I looked at it both ways and decided this way was more informative. Yesterday’s table includes both prepared data and tournament games.

The enemy strategies listed in the form “HeavyRush -> SafeExpand” are the initially predicted and the later recognized enemy play, as explained by MicroDK in a comment. When they’re the same, the prediction was correct.

Stardust always plays the same strategy, so it’s no wonder that Microwave was able to predict it. Not that it helped. 3HatchMuta was tried repeatedly because it scored some wins in training.

PurpleWave opened with 2 gate most games. Microwave was able to predict it, but as we saw in UAlbertaBot’s table, the zealots are a Microwave weakness and PurpleWave was able to exploit it. Nevertheless, Microwave was no pushover. The more successful zerg tries were zergling openings, especially variants of the Styx build (9PoolHatchGasSpeed).

BananaBrain is not predictable, and Microwave could not predict its play. Again, the more successful zerg builds were zergling openings.

Microwave was moderately successful in predicting Dragon’s play, because Dragon tends to stick with a successful strategy as long as it remains successful. Look at that mix of zerg openings! 4 pool, hydra builds, and mutalisk builds.

Microwave tried a lot of stuff versus McRave—three hatch before pool hydralisk opening in ZvZ? And it worked sometimes? I should try to find some of those games.

Microwave recognizes turtle builds in most games. That will be Steamhammer’s OverpoolTurtle opening, which builds as many sunkens at it can afford (2) without delaying mutalisks. It’s tough for bots to handle, because the build is safe on the ground while giving nothing away in the air. Microwave mainly preferred speed zergling openings in response, taking advantage of its superior zergling-on-zergling micro (which is not really a difference in micro as much as in engagement skills).

It looks like ZZZKBot played its 4 pool about 2/3 of the time, and the rest of the time did something that Microwave could not recognize. But no matter, Microwave played overpool nearly all the time, fast enough to stop the rush and, in Microwave’s hands, flexible enough to counter ZZZKBot’s other builds.

Compare this to UAlbertaBot’s table. Microwave did not do perfectly against any UAlbertaBot race, and suffered badly against the zealot rush. Microwave had neither a universal build that works against all UAlbertaBot plays (which is how Steamhammer succeeded against UAlbertaBot), nor was it able to adapt its build well enough to counter what it saw (compare ZZZKBot above). Still, it found that 4 pool and 5 pool were not bad! Fight fire with fire.

Microwave did not understand how to recognize EggBot’s cannon play, but it knew from training how to win.

I’ll cover Microwave over two days because it writes two files for each opponent, a “results” file giving wins/losses for each strategy and a “history” file of more detailed game records. Each summary is bulky in itself, and I don’t want to pile them up. The history file has all the information in the results file and more. In fact, a quick look at Microwave’s code says that it no longer reads the results file at all, but reconstructs its contents from the history file each game. But different presentations of the data have value in themselves; this view makes it easy to read across the columns and see where a given opening was effective.

Today is the results file, the table of strategies versus each opponent. Wow, that’s a lot of opening builds! I count 73, less than half as many as Steamhammer but still too large a number to explore in a tournament of 150 rounds. I think only bots with combinatorial strategies have more. The numbers include not only games played during the tournament, but also Microwave’s prepared data for each opponent, so they add up to more than 150 games versus each opponent. You can compare the overall win rates per opponent to see which ones Microwave was more successful against in training as opposed to in the tournament—it may indicate whether the opponent was updated for the tournament and became stronger than Microwave expected. In general, for stronger opponents training data overestimated Microwave’s success, while for weaker opponents it was the opposite (that is, the training uncovered mistakes that Microwave could then avoid).

	total	#1 stardust	#2 purplewave	#3 bananabrain	#4 dragon	#5 mcrave	#7 steamhammer	#8 daqin	#9 zzzkbot	#10 ualbertabot	#11 willyt	#12 ecgberht	#13 eggbot
10Hatch9Pool9gas	28-48 37%	0-4 0%	0-12 0%	3-19 14%	-	0-1 0%	4-5 44%	-	-	-	21-7 75%	-	-
10HatchMain9Pool9Gas	2-8 20%	0-1 0%	-	0-1 0%	-	1-3 25%	1-3 25%	-	-	-	-	-	-
10HatchTurtleHydra	0-3 0%	-	-	-	0-1 0%	0-1 0%	0-1 0%	-	-	-	-	-	-
11HatchTurtleHydra	0-11 0%	-	0-10 0%	0-1 0%	-	-	-	-	-	-	-	-	-
11HatchTurtleLurker	11-17 39%	-	0-1 0%	-	0-2 0%	0-1 0%	0-1 0%	-	-	-	11-12 48%	-	-
11HatchTurtleMuta	4-15 21%	-	0-7 0%	0-2 0%	-	-	0-1 0%	-	-	-	4-5 44%	-	-
12Hatch	0-3 0%	-	-	0-2 0%	-	0-1 0%	-	-	-	-	-	-	-
12HatchMain	0-4 0%	0-2 0%	0-1 0%	-	-	-	0-1 0%	-	-	-	-	-	-
12Pool	35-51 41%	0-1 0%	-	-	-	9-17 35%	26-33 44%	-	-	-	-	-	-
12PoolMain	25-34 42%	0-1 0%	-	22-21 51%	0-1 0%	0-2 0%	2-7 22%	-	-	-	1-2 33%	-	-
12PoolMuta	7-20 26%	0-2 0%	2-9 18%	0-1 0%	-	-	-	-	-	-	5-8 38%	-	-
1HatchMuta_Sparkle	56-9 86%	-	-	0-1 0%	-	-	-	56-6 90%	-	0-2 0%	-	-	-
2HatchHydra	161-24 87%	0-1 0%	0-2 0%	-	-	-	-	0-1 0%	-	-	-	161-20 89%	-
2HatchLurker	0-8 0%	0-1 0%	0-1 0%	0-2 0%	-	0-1 0%	0-2 0%	-	-	-	-	0-1 0%	-
2HatchLurkerAllIn	0-2 0%	0-1 0%	0-1 0%	-	-	-	-	-	-	-	-	-	-
2HatchMuta	74-59 56%	1-14 7%	0-1 0%	3-9 25%	60-33 65%	0-1 0%	-	-	-	-	-	-	10-1 91%
2HatchMuta_Sparkle	0-1 0%	-	-	-	-	-	-	-	-	-	0-1 0%	-	-
3Hatch	0-5 0%	0-3 0%	-	-	-	0-2 0%	-	-	-	-	-	-	-
3HatchExpo	0-3 0%	0-2 0%	-	-	-	-	-	-	-	-	0-1 0%	-	-
3HatchHydra	0-2 0%	0-1 0%	-	-	-	-	-	-	-	-	0-1 0%	-	-
3HatchHydraBust	0-14 0%	0-7 0%	0-2 0%	0-2 0%	-	0-1 0%	0-1 0%	0-1 0%	-	-	-	-	-
3HatchHydraExpo	1-5 17%	-	-	-	1-2 33%	-	0-2 0%	-	-	0-1 0%	-	-	-
3HatchHydra_BHG	0-4 0%	0-1 0%	0-1 0%	0-1 0%	-	0-1 0%	-	-	-	-	-	-	-
3HatchLingBust	36-41 47%	2-20 9%	-	2-6 25%	-	-	-	32-15 68%	-	-	-	-	-
3HatchLurker	0-4 0%	-	-	-	-	0-1 0%	-	-	-	-	0-3 0%	-	-
3HatchMuta	90-106 46%	7-58 11%	3-9 25%	3-19 14%	-	0-1 0%	-	72-14 84%	-	-	5-5 50%	-	-
3HatchMutaExpo	48-64 43%	0-1 0%	1-25 4%	0-1 0%	32-22 59%	0-1 0%	-	9-7 56%	-	-	6-7 46%	-	-
3HatchPoolHydra	1-24 4%	1-15 6%	0-2 0%	-	-	0-2 0%	0-2 0%	0-3 0%	-	-	-	-	-
3HatchPoolHydraExpo	3-12 20%	0-1 0%	-	0-1 0%	-	3-9 25%	-	0-1 0%	-	-	-	-	-
4HatchBeforeGas	0-10 0%	0-5 0%	0-1 0%	0-3 0%	-	0-1 0%	-	-	-	-	-	-	-
4HatchPoolHydra	4-25 14%	0-2 0%	0-1 0%	2-15 12%	2-6 25%	-	0-1 0%	-	-	-	-	-	-
4PoolHard	3-13 19%	0-1 0%	1-7 12%	2-4 33%	-	-	-	-	-	-	0-1 0%	-	-
4PoolSoft	61-44 58%	0-3 0%	0-2 0%	0-2 0%	7-14 33%	0-3 0%	-	-	-	54-17 76%	0-3 0%	-	-
5HatchPoolHydra	5-28 15%	2-26 7%	-	-	3-2 60%	-	-	-	-	-	-	-	-
5Pool	7-17 29%	0-2 0%	-	-	-	0-1 0%	3-10 23%	0-1 0%	-	4-3 57%	-	-	-
5PoolSpeed	29-18 62%	-	-	-	0-1 0%	0-1 0%	1-2 33%	-	-	28-14 67%	-	-	-
6Pool	0-6 0%	0-2 0%	0-1 0%	0-2 0%	-	-	0-1 0%	-	-	-	-	-	-
6PoolSpeed	4-14 22%	0-4 0%	3-7 30%	-	-	1-3 25%	-	-	-	-	-	-	-
7Pool	0-6 0%	0-1 0%	0-3 0%	0-1 0%	-	-	0-1 0%	-	-	-	-	-	-
7PoolHydraLingRush7D	0-5 0%	-	0-1 0%	-	0-1 0%	0-1 0%	0-1 0%	-	-	0-1 0%	-	-	-
7PoolHydraRush7D	0-2 0%	-	-	-	-	-	-	0-1 0%	-	0-1 0%	-	-	-
8Pool	0-2 0%	-	-	0-1 0%	-	0-1 0%	-	-	-	-	-	-	-
8PoolHydraRush8D	5-8 38%	-	0-2 0%	0-1 0%	-	-	-	-	-	5-5 50%	-	-	-
9Hatch9Pool9Gas	10-12 45%	0-1 0%	0-6 0%	0-1 0%	-	-	0-1 0%	-	-	-	10-3 77%	-	-
9HatchMain8Pool8Gas	2-8 20%	0-1 0%	2-6 25%	-	-	0-1 0%	-	-	-	-	-	-	-
9HatchTurtleHydra	0-3 0%	0-1 0%	0-1 0%	-	-	-	0-1 0%	-	-	-	-	-	-
9Pool	183-11 94%	0-4 0%	0-3 0%	-	-	0-1 0%	-	0-1 0%	-	-	0-1 0%	0-1 0%	183-0 100%
9PoolExpo	31-9 78%	-	-	-	-	-	-	-	-	-	31-9 78%	-	-
9PoolGasHatchSpeed7D	18-19 49%	0-1 0%	18-18 50%	-	-	-	-	-	-	-	-	-	-
9PoolGasHatchSpeed8D	21-29 42%	0-4 0%	-	12-6 67%	-	0-2 0%	0-1 0%	0-3 0%	-	1-1 50%	8-12 40%	-	-
9PoolHatch	0-3 0%	-	-	-	-	-	0-2 0%	-	0-1 0%	-	-	-	-
9PoolHatchGasSpeed7D	11-8 58%	0-1 0%	11-6 65%	0-1 0%	-	-	-	-	-	-	-	-	-
9PoolHatchGasSpeed8D	36-50 42%	1-16 6%	5-6 45%	30-24 56%	0-1 0%	0-2 0%	0-1 0%	-	-	-	-	-	-
9PoolHydra	0-3 0%	0-1 0%	-	-	-	-	-	0-1 0%	-	-	0-1 0%	-	-
9PoolLurker	15-14 52%	0-2 0%	-	-	-	-	-	-	-	-	6-8 43%	9-4 69%	-
9PoolSpeed	68-71 49%	0-3 0%	1-4 20%	0-4 0%	-	13-16 45%	24-21 53%	0-2 0%	-	-	30-21 59%	-	-
9PoolSpeedLing	2-22 8%	0-3 0%	0-3 0%	2-7 22%	-	0-3 0%	0-2 0%	0-3 0%	-	-	0-1 0%	-	-
9PoolSunkHatch	6-12 33%	0-2 0%	0-1 0%	0-1 0%	0-1 0%	-	-	-	-	5-4 56%	1-3 25%	-	-
9PoolSunken	9-15 38%	0-1 0%	0-1 0%	-	-	0-1 0%	2-10 17%	-	-	-	7-2 78%	-	-
Overpool	0-4 0%	0-1 0%	-	0-1 0%	0-1 0%	0-1 0%	-	-	-	-	-	-	-
OverpoolLurker	0-4 0%	0-1 0%	-	-	0-1 0%	-	-	0-1 0%	-	-	0-1 0%	-	-
OverpoolSpeed	187-60 76%	0-1 0%	1-6 14%	0-3 0%	-	27-37 42%	1-3 25%	-	158-9 95%	-	0-1 0%	-	-
OverpoolSunken	8-7 53%	-	-	-	-	-	-	-	-	8-7 53%	-	-	-
OverpoolTurtle	1-13 7%	0-1 0%	1-11 8%	-	0-1 0%	-	-	-	-	-	-	-	-
ZvP_10Hatch9Pool	17-30 36%	0-5 0%	0-8 0%	2-5 29%	-	-	-	-	-	15-12 56%	-	-	-
ZvP_11Hatch10Pool	2-12 14%	0-2 0%	0-7 0%	0-1 0%	-	-	2-2 50%	-	-	-	-	-	-
ZvP_2HatchHydra	2-28 7%	0-3 0%	0-9 0%	0-4 0%	-	0-2 0%	2-10 17%	-	-	-	-	-	-
ZvP_9Hatch9Pool	13-34 28%	0-2 0%	0-10 0%	0-1 0%	-	13-18 42%	0-1 0%	0-2 0%	-	-	-	-	-
ZvZ_Overgas11Pool	30-24 56%	0-3 0%	-	-	-	-	13-15 46%	-	17-5 77%	-	0-1 0%	-	-
ZvZ_Overgas9Pool	0-5 0%	0-2 0%	-	0-1 0%	-	-	0-2 0%	-	-	-	-	-	-
ZvZ_Overpool11Gas	5-21 19%	0-2 0%	-	0-2 0%	-	5-13 28%	0-3 0%	-	-	0-1 0%	-	-	-
ZvZ_Overpool9Gas	2-13 13%	0-2 0%	0-1 0%	-	-	1-7 12%	1-3 25%	-	-	-	-	-	-
ZvZ_OverpoolTurtle	0-1 0%	0-1 0%	-	-	-	-	-	-	-	-	-	-	-
total	-  51%	14-250  5%	49-216  18%	83-180  32%	105-90  54%	73-161  31%	82-153  35%	169-63  73%	175-15  92%	120-69  63%	146-120  55%	170-26  87%	193-1  99%

Microwave explored widely against top opponents, and concentrated efficiently on a few winning openings against weaker ones. On the other hand, although there is a flag in the configuration file named PlayGoodStrategiesFirst (turned on), Microwave seems to have little idea which strategies are most likely to work. Versus DaQin, 1 hatch mutalisk and 3 hatch mutalisk are successful, but the most natural 2 hatch muta is never tried. Of course that’s a widespread weakness among bots.

The 3 hatch muta strategies were relatively successful overall. That’s interesting.

Though UAlbertaBot has been surpassed over the years and become a low-end bot, we can still gain insight from its experience. The table summarizes the contents of its learning files. Last year this table had the bots down the left and the strategies across the top, but this year I turned it on its side—I am looking ahead to the table for Microwave, which has many strategies.

Some of the numbers here are slightly different from those in the official crosstable, because of games where UAlbertaBot did not record a result (no doubt due to crashes).

	total	#1 stardust	#2 purplewave	#3 bananabrain	#4 dragon	#5 mcrave	#6 microwave	#7 steamhammer	#8 daqin	#9 zzzkbot	#11 willyt	#12 ecgberht	#13 eggbot
4RaxMarines	33-120 22%	2-18 10%	1-6 14%	0-13 0%	1-8 11%	7-12 37%	3-14 18%	0-9 0%	0-6 0%	0-10 0%	17-19 47%	2-5 29%	-
MarineRush	66-140 32%	0-10 0%	0-5 0%	1-18 5%	6-16 27%	8-12 40%	4-12 25%	4-23 15%	0-5 0%	0-10 0%	0-5 0%	6-9 40%	37-15 71%
TankPush	17-105 14%	0-10 0%	5-18 22%	0-12 0%	0-5 0%	0-3 0%	0-5 0%	0-9 0%	8-23 26%	0-9 0%	4-8 33%	0-3 0%	-
VultureRush	17-82 17%	0-10 0%	0-5 0%	0-12 0%	0-5 0%	0-3 0%	0-5 0%	0-9 0%	0-5 0%	0-9 0%	0-3 0%	17-16 52%	-
DTRush	19-119 14%	1-22 4%	0-13 0%	1-13 7%	11-26 30%	0-16 0%	-	0-14 0%	6-13 32%	-	0-2 0%	-	-
DragoonRush	13-117 10%	0-16 0%	0-13 0%	0-9 0%	0-12 0%	2-28 7%	-	1-20 5%	10-17 37%	-	0-2 0%	-	-
ZealotRush	185-170 52%	0-15 0%	2-24 8%	4-20 17%	0-12 0%	0-16 0%	37-13 74%	1-20 5%	0-5 0%	48-11 81%	25-18 58%	28-15 65%	40-1 98%
2HatchHydra	9-83 10%	0-12 0%	2-21 9%	1-6 14%	6-13 32%	0-11 0%	0-6 0%	0-8 0%	0-2 0%	0-1 0%	0-3 0%	-	-
3HatchMuta	0-63 0%	0-12 0%	0-12 0%	0-4 0%	0-4 0%	0-11 0%	0-6 0%	0-8 0%	0-2 0%	0-1 0%	0-3 0%	-	-
3HatchScourge	0-59 0%	0-11 0%	0-11 0%	0-4 0%	0-4 0%	0-10 0%	0-6 0%	0-7 0%	0-2 0%	0-1 0%	0-3 0%	-	-
ZerglingRush	195-168 54%	0-11 0%	0-11 0%	11-21 34%	5-11 31%	0-10 0%	14-23 38%	1-11 8%	22-23 49%	29-18 62%	21-17 55%	39-10 80%	53-2 96%
total	-  31%	3-147  2%	10-139  7%	18-132  12%	29-116  20%	17-132  11%	58-90  39%	7-138  5%	46-103  31%	77-70  52%	67-83  45%	92-58  61%	130-18  88%

Random UAlbertaBot starts off with its default strategies of marine rush, zealot rush, or zergling rush, and tries alternatives only if the strategy scores poorly. The table shows that the default strategies chosen years ago are still the best choices. The zealot rush even scored well against #6 Microwave. Also constant over the years is that the 3 hatch scourge build, which was designed to counter the carrier bot XIMP, has no other use; UAlbertaBot would have done better without it.

It’s curious that UAlbertaBot’s overall weakest race is terran, but that its terran scored best against many stronger opponents: Terran was UAlbertaBot’s happiest roll versus #7 Steamhammer, #5 McRave, #2 PurpleWave, and #1 Stardust. But these stronger opponents allowed few wins. #1 Stardust (2%), #7 Steamhammer (5%), and #2 PurpleWave (7%) shut down UAlbertaBot hard.

If your bot is ranked above UAlbertaBot, then pink or blue boxes suggest weaknesses that you might benefit from working on. If a weaker bot beats you this way, presumably a stronger one can too. UAlbertaBot benefits from its random race and the big differences between its strategies, so maybe something went wrong in your scouting or reactions. #6 Microwave had trouble with zealots, #5 McRave had trouble with marines, and #4 Dragon had some trouble with 4 different rushes.

Steamhammer 3.2.20 is uploaded to SSCAIT, source code to follow shortly. It has minor bug fixes and tweaks, nothing earthshaking. I turned on the DrawWorkerInfo debug flag to add fun. Here’s what’s new.

squad targeting

• A squad that is empty or otherwise cannot attack is given a default target. This “amazing” optimization may save a few hundred microseconds of computation per game, though I think it’s less. I made this change after investigating what I suspected was a serious bug that gave squads bad targets. The cause turned out to be a race between squad updating and squad targeting that went wrong a few times per game and affected a squad for only one frame, and only when the squad wasn’t able to take action anyway—the bug actually had no bad effects. This change papers over the bug without addressing the unimportant race condition.

infrastructure

• go post worker now works on frame 0. The feature worked when I originally released it, but a change I made to fix a problem for AIIDE (reordering manager calls to prevent double commanding) caused go post worker to fail when issued at the very beginning of the build order. I fixed it by removing the worker job type Default; a worker that has not yet been given a job is now Idle, as it logically should be. I like it when I can fix a bug by cutting out unnecessary parts. The effect of the bug was to break one protoss build, Proxy2Gate, that sends out a worker first thing.

• MapGrid::getLeastExplored() no longer accepts a zone argument. It’s not needed since getLeastExploredNear() was added.

construction

• The building manager correctly counts the number of workers sent to construct a building. If the first worker dies and it has to send another, the count goes to 2, and so on. It used to often miscount. This is the most important fix here, and solves the zerg macro hatch conversion bug. It should also solve misbehaviors by terran and protoss.

terran

• It was rare but possible for a terran building to be abandoned while incomplete, lost from the building queue so that it was never canceled, permanently wasting resources. Fixed.

• When going bio, get gas and academy slightly earlier.

zerg

• Assign an overlord to look over the next base that Steamhammer intends to take. That will let it spot any spider mines or other obstacles to taking the base. The overlord usually arrives well before any drone is sent. The overlord stays home if the enemy has flying overlord hunters out on the map, though.

• Adjust desired drone counts so that Steamhammer is willing to make more queens. For some reason, Steamhammer was rarely making more than 1 queen. With this change, it often makes 2 when the enemy units and queen research justify it (many tanks means more queens for broodling). It’s configured to make up to 6 queens, so 2 is still low, but the issue is not worth a close look right now.

openings

• Over10PoolHydra hydra bust added. This was inspired by a CUBOT build that I saw drive BananaBrain to its knees (CUBOT broke the protoss natural, continued its attack into the main and killed nearly all probes, then lost despite a winning advantage when already queued dragoons popped out of the gates). Steamhammer’s version is entirely different in detail and I think stronger.

Next should be Steamhammer 3.3, finally ported to BWAPI 4.4.0. Tomorrow I’ll start looking at what different bots learned from their experience in AIIDE. Meanwhile, the AIIDE unknown maps competition must be running now, and the results may be out this weekend or thereabouts. I’m eager to see what maps came up this time.

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. Standard disclaimers apply: 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	Nothing. Stardust relies on its great execution.
2	PurpleWave	The learning files have a sequence of PurpleWave’s strategy choices followed by a sequence of “fingerprinted” enemy strategies. (PurpleWave also has specific preparation for its opponents, but that’s in code rather than data.) There are also debug logs that show some decisions, but are probably only for the author.
3	BananaBrain	The learning files look just like last year’s: 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 6 opponents, with the largest file by far for Stardust. Maybe if you have pegged your opponent as having a narrow range of adaptation, you don’t have to leave room for surprises.
4	Dragon	Very simple game records with strategy and game result, like "siege expand" won.
5	McRave	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 (now working correctly, unlike last year) 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. I want to look more closely at the game records and see how they are used (maybe they are only logs for the author).
6	Microwave	Result and history files for each opponent that look similar to last year’s. The result files count wins and losses for each Microwave strategy, and no longer limit the counts to 10—apparently Microwave no longer deliberately forgets history. The history files have a one-line record of data about each game and look the same as last year. Also pre-learned history files for all 12 opponents.
7	Steamhammer	Steamhammer’s learning file format is documented here.
8	DaQin	Carried over from last year. Learning files straight from its parent Locutus (very similar to the old format Steamhammer files). There is no visible pre-learned data (in a quick check I also found no opponent-specific code).
9	ZZZKBot	Learning files for each opponent that look the same as last year, with detailed but hard-to-interpret information about each game.
10	UAlbertaBot	Carried over from past years. For each opponent, a file listing strategies with win and loss counts for each.
11	WillyT	A single log file with 150 lines apparently giving data for 150 games against various opponents. Each line looks like 20201009,Ecgberht,T,01,0. The items look like date, opponent, opponent race, a number 01 02 or 03, and win/loss. There were 150 rounds in the tournament, so maybe this is a log of one game per round—the dates seem to back that up, but if so, how is the single game chosen? Is it the last one played? This is either broken, or else it is doing something I can’t fathom.
12	Ecgberht	Two files for each opponent, named like Dragon_Terran.json and Dragon_Terran-History.json. The plain file counts wins and losses of each of Ecgberht’s strategies separately for each map size (number of starting locations, 2 3 or 4). (The map size breakdown is similar to AIUR’s.) There is also an overall win/loss count, plus flags named naughty and defendHarass. Of all bots in the tournament, only ZZZKBot is flagged naughty, so maybe it means the opponent likes fast rushes. defendHarass tells whether the opponent defends its workers if Ecgberht’s scouting SCV attacks them (that way it can exploit weak opponents without risking its SCV against prepared ones). The history file is a list of game records, giving opponent name, opponent race, game outcome, Ecgberht’s strategy, the map, and the opponent’s recognized strategy (which is often Unknown).
13	EggBot	Nothing. EggBot is the only entrant other than Stardust to record no data.

In recent years, nearly all top bots have relied on opening learning to adapt to their opponents. The strongest bot without learning was Iron, which came in #1 in AIIDE 2016 and slipped down the ranks until it fell to #8 in AIIDE 2019, scoring under 50%. Stardust is the only high finisher since then to get by without. Stardust plays with a restricted set of units, only zealots and dragoons with observers as needed. On the one hand, that shows the value of specializing and becoming extremely skilled at the most important aspects of the game (the opposite of Steamhammer’s development strategy). On the other hand, it points out how much headroom all bots have to improve.

I’ve found that the bland average “how well did I do on this map?” often obscures big variations in “how well did I do versus each opponent on this map?” The variations mean something about strategy execution or strategy selection. Unfortunately, with only this information it’s usually impossible to guess what the meaning is, so I don’t have many comments. In general, I think you have to look at games to see how two bots’ strategies interact on a given map. These tables can help you choose which games to look at.

First, here’s the overall table with the averages. It’s identical in content to the third table in the official results. With 1800 games per player and 10 maps, each cell represents 180 games (sometimes slightly fewer due to missing games). I also have a writeup of the maps.

	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	93.22%	96%	92%	96%	93%	88%	94%	92%	94%	94%	92%
purplewave	79.44%	79%	82%	76%	80%	81%	76%	79%	82%	80%	81%
bananabrain	69.61%	65%	73%	69%	75%	71%	71%	68%	71%	65%	69%
dragon	62.38%	58%	63%	67%	63%	57%	69%	65%	59%	61%	60%
mcrave	57.22%	62%	54%	57%	59%	58%	54%	55%	61%	56%	56%
microwave	54.47%	58%	59%	47%	54%	54%	52%	52%	54%	61%	53%
steamhammer	54.00%	55%	53%	58%	52%	56%	50%	60%	53%	52%	51%
daqin	50.14%	53%	48%	51%	51%	51%	56%	51%	49%	42%	49%
zzzkbot	39.89%	35%	37%	32%	40%	39%	41%	39%	38%	53%	45%
ualbertabot	31.14%	33%	30%	35%	32%	37%	24%	28%	34%	23%	34%
willyt	29.44%	19%	26%	32%	22%	29%	35%	38%	33%	33%	28%
ecgberht	24.28%	28%	28%	27%	21%	24%	23%	21%	18%	23%	29%
eggbot	4.72%	9%	4%	4%	7%	5%	4%	2%	4%	4%	3%

Now the breakdown tables. For example, in the next table, Stardust scored 80% against PurpleWave on the map Destination. Each of these cells represents only 15 games (minus missing games), so small differences are not significant.

stardust	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
purplewave	83%	80%	80%	93%	93%	73%	100%	80%	87%	73%	73%
bananabrain	62%	87%	53%	73%	33%	47%	67%	60%	53%	87%	60%
dragon	93%	100%	80%	93%	100%	93%	87%	93%	93%	100%	87%
mcrave	98%	93%	100%	93%	100%	100%	93%	100%	100%	100%	100%
microwave	99%	100%	100%	100%	100%	93%	100%	100%	100%	100%	100%
steamhammer	98%	100%	100%	100%	100%	87%	100%	93%	100%	100%	100%
daqin	93%	87%	100%	100%	93%	87%	93%	87%	100%	87%	93%
zzzkbot	99%	100%	100%	100%	100%	100%	100%	93%	100%	100%	100%
ualbertabot	98%	100%	100%	100%	100%	80%	100%	100%	100%	100%	100%
willyt	99%	100%	100%	100%	100%	100%	100%	100%	100%	93%	100%
ecgberht	99%	100%	100%	100%	100%	100%	100%	100%	100%	93%	93%
eggbot	97%	100%	87%	100%	100%	100%	93%	100%	100%	93%	100%
overall	93.22%	96%	92%	96%	93%	88%	94%	92%	94%	94%	92%

Stardust did about equally well on every map against PurpleWave, but not so against BananaBrain. It looks as though it’s due to BananaBrain’s play—see the BananaBrain table below. Stardust thoroughly suppressed other players, so that’s as much as we can say.

purplewave	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	17%	20%	20%	7%	7%	27%	0%	20%	13%	27%	27%
bananabrain	63%	60%	87%	47%	60%	60%	73%	67%	60%	67%	47%
dragon	45%	53%	47%	47%	40%	53%	7%	27%	80%	40%	53%
mcrave	95%	100%	87%	87%	100%	100%	100%	100%	80%	93%	100%
microwave	71%	53%	67%	67%	80%	67%	73%	67%	87%	67%	80%
steamhammer	95%	93%	100%	93%	93%	100%	93%	100%	87%	93%	100%
daqin	87%	80%	87%	87%	80%	87%	93%	73%	100%	93%	93%
zzzkbot	91%	93%	93%	100%	100%	73%	87%	100%	93%	93%	73%
ualbertabot	93%	93%	93%	80%	100%	100%	93%	100%	93%	87%	93%
willyt	98%	100%	100%	93%	100%	100%	93%	100%	93%	100%	100%
ecgberht	99%	100%	100%	100%	100%	100%	93%	100%	100%	100%	100%
eggbot	100%	100%	100%	100%	100%	100%	100%	100%	100%	100%	100%
overall	79.44%	79%	82%	76%	80%	81%	76%	79%	82%	80%	81%

bananabrain	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	38%	13%	47%	27%	67%	53%	33%	40%	47%	13%	40%
purplewave	37%	40%	13%	53%	40%	40%	27%	33%	40%	33%	53%
dragon	45%	40%	47%	40%	40%	60%	47%	27%	33%	60%	53%
mcrave	55%	47%	73%	40%	60%	53%	60%	40%	53%	60%	60%
microwave	61%	60%	87%	60%	67%	53%	53%	67%	67%	33%	67%
steamhammer	71%	60%	80%	60%	87%	53%	80%	73%	73%	67%	73%
daqin	66%	60%	67%	53%	67%	67%	53%	67%	80%	80%	67%
zzzkbot	94%	100%	100%	100%	100%	87%	100%	100%	93%	87%	73%
ualbertabot	88%	93%	80%	100%	80%	87%	100%	87%	80%	100%	73%
willyt	84%	73%	100%	93%	93%	93%	93%	80%	80%	60%	73%
ecgberht	99%	93%	93%	100%	100%	100%	100%	100%	100%	100%	100%
eggbot	98%	100%	93%	100%	100%	100%	100%	100%	100%	87%	100%
overall	69.61%	65%	73%	69%	75%	71%	71%	68%	71%	65%	69%

BananaBrain shows an irregular pattern against its strongest opposition, but seems to have more consistent map preferences against the middle ranks. I don’t know what it means, but it’s interesting.

dragon	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	7%	0%	20%	7%	0%	7%	13%	7%	7%	0%	13%
purplewave	55%	47%	53%	53%	60%	47%	93%	73%	20%	60%	47%
bananabrain	55%	60%	53%	60%	60%	40%	53%	73%	67%	40%	47%
mcrave	79%	80%	73%	93%	73%	80%	80%	73%	47%	93%	93%
microwave	57%	27%	40%	80%	53%	47%	80%	60%	60%	60%	60%
steamhammer	30%	20%	27%	33%	33%	20%	47%	20%	40%	20%	40%
daqin	53%	40%	53%	27%	33%	60%	53%	80%	53%	60%	73%
zzzkbot	47%	47%	67%	73%	67%	40%	40%	33%	67%	20%	20%
ualbertabot	80%	80%	92%	87%	80%	60%	87%	80%	71%	87%	80%
willyt	93%	100%	93%	100%	100%	93%	93%	87%	93%	93%	80%
ecgberht	94%	100%	100%	100%	100%	100%	100%	93%	80%	100%	67%
eggbot	97%	100%	93%	93%	100%	93%	93%	100%	100%	100%	100%
overall	62.38%	58%	63%	67%	63%	57%	69%	65%	59%	61%	60%

Dragon is strangely inconsistent. Does it learn strategy independently for each map? Does it do map analysis to adapt its play to the map, but there are flaws? Look at the PurpleWave row: Dragon crushed it on Circuit Breaker and fell down on Fighting Spirit.

mcrave	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	2%	7%	0%	7%	0%	0%	7%	0%	0%	0%	0%
purplewave	5%	0%	13%	13%	0%	0%	0%	0%	20%	7%	0%
bananabrain	45%	53%	27%	60%	40%	47%	40%	60%	47%	40%	40%
dragon	21%	20%	27%	7%	27%	20%	20%	27%	53%	7%	7%
microwave	77%	87%	47%	73%	80%	87%	80%	73%	93%	80%	73%
steamhammer	57%	53%	33%	60%	47%	67%	80%	60%	53%	53%	67%
daqin	65%	87%	80%	87%	87%	47%	27%	33%	60%	80%	60%
zzzkbot	83%	93%	87%	87%	87%	93%	73%	73%	87%	60%	87%
ualbertabot	89%	80%	100%	87%	93%	93%	93%	93%	87%	80%	80%
willyt	63%	80%	53%	40%	80%	67%	47%	60%	47%	73%	80%
ecgberht	80%	80%	87%	60%	73%	80%	80%	80%	87%	93%	80%
eggbot	99%	100%	100%	100%	93%	100%	100%	100%	100%	100%	100%
overall	57.22%	62%	54%	57%	59%	58%	54%	55%	61%	56%	56%

microwave	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	1%	0%	0%	0%	0%	7%	0%	0%	0%	0%	0%
purplewave	29%	47%	33%	33%	20%	33%	27%	33%	13%	33%	20%
bananabrain	39%	40%	13%	40%	33%	47%	47%	33%	33%	67%	33%
dragon	43%	73%	60%	20%	47%	53%	20%	40%	40%	40%	40%
mcrave	23%	13%	53%	27%	20%	13%	20%	27%	7%	20%	27%
steamhammer	29%	53%	47%	13%	33%	13%	33%	13%	27%	33%	20%
daqin	83%	87%	100%	73%	80%	93%	53%	87%	80%	100%	80%
zzzkbot	93%	93%	93%	100%	93%	93%	93%	100%	100%	73%	93%
ualbertabot	61%	47%	60%	50%	53%	53%	87%	60%	60%	80%	60%
willyt	65%	87%	73%	27%	73%	47%	53%	47%	87%	87%	67%
ecgberht	88%	60%	80%	80%	100%	93%	93%	80%	100%	100%	93%
eggbot	100%	100%	100%	100%	100%	100%	100%	100%	100%	100%	100%
overall	54.47%	58%	59%	47%	54%	54%	52%	52%	54%	61%	53%

steamhammer	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	2%	0%	0%	0%	0%	13%	0%	7%	0%	0%	0%
purplewave	5%	7%	0%	7%	7%	0%	7%	0%	13%	7%	0%
bananabrain	29%	40%	20%	40%	13%	47%	20%	27%	27%	33%	27%
dragon	70%	80%	73%	67%	67%	80%	53%	80%	60%	80%	60%
mcrave	43%	47%	67%	40%	53%	33%	20%	40%	47%	47%	33%
microwave	71%	47%	53%	87%	67%	87%	67%	87%	73%	67%	80%
daqin	22%	40%	20%	20%	13%	13%	27%	40%	7%	20%	20%
zzzkbot	75%	60%	73%	93%	67%	93%	80%	93%	87%	47%	53%
ualbertabot	95%	80%	100%	93%	100%	93%	87%	100%	100%	100%	100%
willyt	55%	60%	67%	60%	60%	53%	60%	60%	40%	33%	60%
ecgberht	83%	100%	60%	87%	87%	60%	93%	93%	87%	93%	73%
eggbot	97%	100%	100%	100%	93%	100%	87%	93%	100%	100%	100%
overall	54.00%	55%	53%	58%	52%	56%	50%	60%	53%	52%	51%

daqin	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	7%	13%	0%	0%	7%	13%	7%	13%	0%	13%	7%
purplewave	13%	20%	13%	13%	20%	13%	7%	27%	0%	7%	7%
bananabrain	34%	40%	33%	47%	33%	33%	47%	33%	20%	20%	33%
dragon	47%	60%	47%	73%	67%	40%	47%	20%	47%	40%	27%
mcrave	35%	13%	20%	13%	13%	53%	73%	67%	40%	20%	40%
microwave	17%	13%	0%	27%	20%	7%	47%	13%	20%	0%	20%
steamhammer	78%	60%	80%	80%	87%	87%	73%	60%	93%	80%	80%
zzzkbot	9%	40%	7%	20%	7%	7%	0%	0%	0%	0%	13%
ualbertabot	69%	73%	87%	53%	67%	60%	73%	73%	73%	71%	60%
willyt	96%	100%	100%	93%	93%	100%	100%	100%	100%	73%	100%
ecgberht	99%	100%	93%	100%	100%	100%	100%	100%	100%	100%	100%
eggbot	97%	100%	100%	87%	100%	100%	100%	100%	100%	87%	100%
overall	50.14%	53%	48%	51%	51%	51%	56%	51%	49%	42%	49%

zzzkbot	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	1%	0%	0%	0%	0%	0%	0%	7%	0%	0%	0%
purplewave	9%	7%	7%	0%	0%	27%	13%	0%	7%	7%	27%
bananabrain	6%	0%	0%	0%	0%	13%	0%	0%	7%	13%	27%
dragon	53%	53%	33%	27%	33%	60%	60%	67%	33%	80%	80%
mcrave	17%	7%	13%	13%	13%	7%	27%	27%	13%	40%	13%
microwave	7%	7%	7%	0%	7%	7%	7%	0%	0%	27%	7%
steamhammer	25%	40%	27%	7%	33%	7%	20%	7%	13%	53%	47%
daqin	91%	60%	93%	80%	93%	93%	100%	100%	100%	100%	87%
ualbertabot	49%	33%	47%	20%	67%	33%	53%	67%	40%	87%	40%
willyt	92%	100%	100%	100%	100%	93%	80%	67%	80%	100%	100%
ecgberht	29%	13%	13%	33%	33%	27%	33%	33%	60%	33%	13%
eggbot	100%	100%	100%	100%	100%	100%	100%	100%	100%	100%	100%
overall	39.89%	35%	37%	32%	40%	39%	41%	39%	38%	53%	45%

ualbertabot	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	2%	0%	0%	0%	0%	20%	0%	0%	0%	0%	0%
purplewave	7%	7%	7%	20%	0%	0%	7%	0%	7%	13%	7%
bananabrain	12%	7%	20%	0%	20%	13%	0%	13%	20%	0%	27%
dragon	20%	20%	8%	13%	20%	40%	13%	20%	29%	13%	20%
mcrave	11%	20%	0%	13%	7%	7%	7%	7%	13%	20%	20%
microwave	39%	53%	40%	50%	47%	47%	13%	40%	40%	20%	40%
steamhammer	5%	20%	0%	7%	0%	7%	13%	0%	0%	0%	0%
daqin	31%	27%	13%	47%	33%	40%	27%	27%	27%	29%	40%
zzzkbot	51%	67%	53%	80%	33%	67%	47%	33%	60%	13%	60%
willyt	45%	27%	73%	47%	73%	47%	27%	33%	47%	27%	47%
ecgberht	61%	60%	47%	67%	60%	60%	53%	73%	73%	47%	73%
eggbot	90%	93%	93%	80%	93%	100%	87%	87%	87%	100%	80%
overall	31.14%	33%	30%	35%	32%	37%	24%	28%	34%	23%	34%

willyt	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	1%	0%	0%	0%	0%	0%	0%	0%	0%	7%	0%
purplewave	2%	0%	0%	7%	0%	0%	7%	0%	7%	0%	0%
bananabrain	16%	27%	0%	7%	7%	7%	7%	20%	20%	40%	27%
dragon	7%	0%	7%	0%	0%	7%	7%	13%	7%	7%	20%
mcrave	37%	20%	47%	60%	20%	33%	53%	40%	53%	27%	20%
microwave	35%	13%	27%	73%	27%	53%	47%	53%	13%	13%	33%
steamhammer	45%	40%	33%	40%	40%	47%	40%	40%	60%	67%	40%
daqin	4%	0%	0%	7%	7%	0%	0%	0%	0%	27%	0%
zzzkbot	8%	0%	0%	0%	0%	7%	20%	33%	20%	0%	0%
ualbertabot	55%	73%	27%	53%	27%	53%	73%	67%	53%	73%	53%
ecgberht	74%	53%	87%	47%	80%	93%	80%	93%	87%	60%	60%
eggbot	69%	0%	80%	93%	53%	47%	87%	100%	73%	80%	80%
overall	29.44%	19%	26%	32%	22%	29%	35%	38%	33%	33%	28%

ecgberht	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	1%	0%	0%	0%	0%	0%	0%	0%	0%	7%	7%
purplewave	1%	0%	0%	0%	0%	0%	7%	0%	0%	0%	0%
bananabrain	1%	7%	7%	0%	0%	0%	0%	0%	0%	0%	0%
dragon	6%	0%	0%	0%	0%	0%	0%	7%	20%	0%	33%
mcrave	20%	20%	13%	40%	27%	20%	20%	20%	13%	7%	20%
microwave	12%	40%	20%	20%	0%	7%	7%	20%	0%	0%	7%
steamhammer	17%	0%	40%	13%	13%	40%	7%	7%	13%	7%	27%
daqin	1%	0%	7%	0%	0%	0%	0%	0%	0%	0%	0%
zzzkbot	71%	87%	87%	67%	67%	73%	67%	67%	40%	67%	87%
ualbertabot	39%	40%	53%	33%	40%	40%	47%	27%	27%	53%	27%
willyt	26%	47%	13%	53%	20%	7%	20%	7%	13%	40%	40%
eggbot	97%	93%	100%	100%	87%	100%	100%	100%	93%	100%	100%
overall	24.28%	28%	28%	27%	21%	24%	23%	21%	18%	23%	29%

eggbot	overall	Destin	Heartb	Polari	Aztec	Longin	Circui	Empire	Fighti	Python	Roadki
stardust	3%	0%	13%	0%	0%	0%	7%	0%	0%	7%	0%
purplewave	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%
bananabrain	2%	0%	7%	0%	0%	0%	0%	0%	0%	13%	0%
dragon	3%	0%	7%	7%	0%	7%	7%	0%	0%	0%	0%
mcrave	1%	0%	0%	0%	7%	0%	0%	0%	0%	0%	0%
microwave	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%
steamhammer	3%	0%	0%	0%	7%	0%	13%	7%	0%	0%	0%
daqin	3%	0%	0%	13%	0%	0%	0%	0%	0%	13%	0%
zzzkbot	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%
ualbertabot	10%	7%	7%	20%	7%	0%	13%	13%	13%	0%	20%
willyt	31%	100%	20%	7%	47%	53%	13%	0%	27%	20%	20%
ecgberht	3%	7%	0%	0%	13%	0%	0%	0%	7%	0%	0%
overall	4.72%	9%	4%	4%	7%	5%	4%	2%	4%	4%	3%

EggBot’s wins against WillyT—the majority of all its wins—seem to have come from very specific situations.

I enjoyed lurking in the AIIDE 2020 stream last night. The top winners were easily predicted: #1 Stardust, #2 PurpleWave, #3 BananaBrain. #4 terran Dragon did well, and #5 McRave playing zerg did great to finish as well as they did in the era of protoss domination. The race pattern continues: Protoss at the top and otherwise scattered randomly down the table, terran split between strong bots near the top and weaker bots near the bottom, and zerg clumped in the middle. Of course that is only a general pattern, every bot is on its own.

I will be doing my usual results analysis, I hope more than my usual. I’m curious about how some of the specific results came about.

Here is my version of the crosstable, computed from the detailed results. It exactly matches the official crosstable, only the presentation is different. A total of 5 games went uncounted due to GAME_STATE_NOT_UPDATED_60S_BOTH_BOTS, all of them involving UAlbertaBot.

#	bot	overall	star	purp	bana	drag	mcra	micr	stea	daqi	zzzk	ualb	will	ecgb	eggb
1	stardust	93.22%		83%	62%	93%	98%	99%	98%	93%	99%	98%	99%	99%	97%
2	purplewave	79.44%	17%		63%	45%	95%	71%	95%	87%	91%	93%	98%	99%	100%
3	bananabrain	69.61%	38%	37%		45%	55%	61%	71%	66%	94%	88%	84%	99%	98%
4	dragon	62.38%	7%	55%	55%		79%	57%	30%	53%	47%	80%	93%	94%	97%
5	mcrave	57.22%	2%	5%	45%	21%		77%	57%	65%	83%	89%	63%	80%	99%
6	microwave	54.47%	1%	29%	39%	43%	23%		29%	83%	93%	61%	65%	88%	100%
7	steamhammer	54.00%	2%	5%	29%	70%	43%	71%		22%	75%	95%	55%	83%	97%
8	daqin	50.14%	7%	13%	34%	47%	35%	17%	78%		9%	69%	96%	99%	97%
9	zzzkbot	39.89%	1%	9%	6%	53%	17%	7%	25%	91%		49%	92%	29%	100%
10	ualbertabot	31.14%	2%	7%	12%	20%	11%	39%	5%	31%	51%		45%	61%	90%
11	willyt	29.44%	1%	2%	16%	7%	37%	35%	45%	4%	8%	55%		74%	69%
12	ecgberht	24.28%	1%	1%	1%	6%	20%	12%	17%	1%	71%	39%	26%		97%
13	eggbot	4.72%	3%	0%	2%	3%	1%	0%	3%	3%	0%	10%	31%	3%

It’s curious that Stardust towered over every opponent—only BananaBrain was able to put up a serious fight—but did not score 100% against any. #4 Dragon upset protoss #2 PurpleWave and #3 BananaBrain, but was upset in turn by zergs #7 Steamhammer and #9 ZZZKBot. That is typical of tscmoo authored bots: They are tuned to do well against the best, and show some weakness against the rest. I’m surprised by UAlbertaBot’s relatively high finish; I expected it to be second to last.

In my original post on the bots registered for AIIDE I separated out 3 new bots, DanDanBot, Randofoo, and Taij. I didn’t mention the other new entrant, EggBot, which hadn’t appeared on the list yet. Of the 4 new bots, only EggBot ended up playing. None of the familiar old names dropped out. Way to go EggBot! In my book it did not finish last, it finished ahead of 3 no-shows, and ahead of everyone who was afraid to sign up at all. You don’t have to have a serious chance in the competition to take the competition seriously; opportunities are to be taken. The only downside is that I can no longer say “Eggie” to mean Ecgberht.

I am of course especially interested in Steamhammer’s results. Its rival Microwave squeaked ahead with 8 extra wins out of the 1800 games. #7 Steamhammer upset #4 Dragon and #6 Microwave by about 70% each, but it was crushed by #8 DaQin, scoring only 22% (where Microwave scored 83%). In my post Steamhammer’s prepared learning data for AIIDE 2020 I said “I also didn’t prepare against DaQin because I didn’t have recent data handy; I could have tried harder, but time was short.” That one omission was my downfall!

The race balance tables are not very interesting, since protoss dominates. And of course there was only one random player, UAlbertaBot. Nevertheless, here they are. The overall race balance:

	overall	vT	vP	vZ	vR
terran	39%		31%	38%	58%
protoss	59%	69%		58%	72%
zerg	51%	62%	42%		73%
random	31%	42%	28%	27%

Each bot’s results by opponent race. I think the table tells more about the opponents grouped by race than about the bots listed on the left.

#	bot	race	overall	vT	vP	vZ	vR
1	stardust	protoss	93.22%	97%	84%	99%	98%
2	purplewave	protoss	79.44%	81%	67%	88%	93%
3	bananabrain	protoss	69.61%	76%	60%	70%	88%
4	dragon	terran	62.38%	94%	54%	53%	80%
5	mcrave	zerg	57.22%	55%	43%	72%	89%
6	microwave	zerg	54.47%	65%	50%	48%	61%
7	steamhammer	zerg	54.00%	70%	31%	63%	95%
8	daqin	protoss	50.14%	81%	38%	35%	69%
9	zzzkbot	zerg	39.89%	58%	41%	16%	49%
10	ualbertabot	random	31.14%	42%	28%	27%	-
11	willyt	terran	29.44%	40%	18%	31%	55%
12	ecgberht	terran	24.28%	16%	20%	30%	39%
13	eggbot	protoss	4.72%	12%	2%	1%	10%

Next: Results by map.