AIIDE 2017 the learning bots
In March 2016 I analyzed which bots learned during the AIIDE 2015 tournament by looking at the data files. Here’s a similar analysis for AIIDE 2017.
I looked at the “write” directory for each bot, to see if it wrote files there, and if so, what the files looked like. Writing data doesn’t mean that the bot actually learned anything—it may not have used the data. Bots not listed in the table did not write anything interesting (maybe log files, nothing more). The table includes 15 bots of the 28 entrants, over half of them.
| # | bot | info |
|---|---|---|
| 1 | ZZZKBot | varied info about each game, including tidbits like the time zone and the processor it was played on |
| 2 | PurpleWave | for each opponent, a log of games with info including the sequence of strategies followed |
| 5 | Microwave | same format as UAlbertaBot (Microwave has more strategies) |
| 6 | CherryPi | opening data for each opponent |
| 9 | Tyr | for each opponent, seems to save info only about the previous game: win or loss, and a flag or two like "FLYERS" or "CANNONS" |
| 11 | AILien | 10 values per opponent: scores for zerg unit types, a few odds and ends like macroHeavyness and supplyISawAir |
| 12 | LetaBot | one file which records a log of games, with opponent, race, map, and 3 numbers per game |
| 14 | UAlbertaBot | a file for each opponent, giving for each strategy the count of wins and losses; learning was turned on this year |
| 15 | Aiur | 91 values per opponent: strategy x map size |
| 17 | Skynet | a file for each opponent, with lines like "build_0_2 14 12" |
| 19 | MegaBot | many files; the important ones seem to be "MegaBot-vs-[bot name].xml" which give scores for each bot MegaBot can imitate: Skynet, NUSBot, Xelnaga |
| 20 | Xelnaga | a file for each opponent with a single number: -1, 0, 2, or 3 |
| 21 | Overkill | many files with neural network data, reinforcement learning data, and opening learning data for each opponent (more than I thought!) |
| 24 | Myscbot | same format as UAlbertaBot, but only 1 strategy was played for each opponent; nothing was learned |
| 25 | HannesBredberg | 2 numbers per opponent, approximately (not exactly) the win and loss counts |
LetaBot seems worth looking into, to see whether its log is learning data and, if so, how it is used. PurpleWave also recorded data essentially as a log, which could be used for a wide range of purposes. And AILien has a unique learning method that I should spend some time on.
UAlbertaBot had learning turned on this year. It has sometimes left learning off because its default strategies were dominant. It’s also notable that Ziabot skipped learning this year. In the past it has learned. Ziabot also finished last.
Next: What AIUR learned.
Trackbacks
No Trackbacks
Comments
Dave Churchill on :
MicroDK on :
LetaBot on :
It might be used for learning, but my bot didn't use learning this CIG and AIIDE.
Jay Scott on :