Posted: Thu May 21, 2020 10:47 pm Post subject: Collecting Game Logs for Datasets
I'm not sure if this has come up elsewhere, but I was wondering whether you (MaBi) are saving the game logs of all played games?
They could be very useful for building a dataset to potentially train AI players, or study statistics.
If you are collecting the game logs, do you mind sharing the approximate sizes of how many games you have recorded?
I have the logs but I don't think they would be much useful: they are basically the same logs you see in the game while playing, so they would be difficult to parse to even extract simple statistics, let alone to train an AI.
Currently I have 11362 logs, out of 11580 total games played on mabiweb (the first couple hundred games were mostly for debugging and I didn't log them). Each log is a text file of ~100kB (it contains other info beside the log), for a total of ~1.2GB, ~120MB after zipping.
I was actually considering running some analysis on Nations games played here. So far, I have written a script that parses the HTML of a finished match and outputs a more readable log format:
The idea is to then parse the output log format to study things like correlation between cards and points. It would take many games for that to be meaningful though.
Thoughts, MaBi?
Certainly, if I could seed that analysis with more than 7 or 30 days worth of matches, that would be preferable.
With regard to AI, I think you'd need at least millions of games to train a general one for a game like this. I think the best you could do with thousands of games is figure out relative values of cards and then use that in a hand-written evaluation function.
The idea is to then parse the output log format to study things like correlation between cards and points. It would take many games for that to be meaningful though.
I had done something similar some time ago, parsing the logs for correlations betweens cards and wins, i.e. counting how many games the player who bought a certain card was the final winner. For example, re-running the script with the current 11000 games:
But I don't really know how significant or useful are this kind of stats.
I don’t know either. For one thing, different cards have different values based not only on player count but also “difficulty level†(prince vs. king, etc.). Also, not every card is in every game, even for the games that use every deck. I’m sure I’d be looking up formulas and scratching my head for a while, and might even give up. I’m not a statistician. It might be interesting to try though.
I was thinking I’d try something like this:
https://boardgamegeek.com/thread/1933554/data-driven-strategy-tips
But Through the Ages has the advantage that you see every card in every civil deck. You won’t buy most of them, but at least they’ll be available. There are also far fewer options for play, just number of players.
You cannot post new topics in this forum You cannot reply to topics in this forum You cannot edit your posts in this forum You cannot delete your posts in this forum You cannot vote in polls in this forum
Forum FAQ
Search
Usergroups
Profile
Log in to check your private messages
Log in
