Training a neural network to predict results of setons games

[Memto]

I'd like to train a neural network that can predict the outcomes of seton's games. Our ranking system is already able to give generic predictions on any map, but it does not take into account the map or the spawn positions of the specific players. Since we have a large database of finished setons games that have been played, I believe it is possible to improve the accuracy. The neural network may turn out to be better than the generic algorithm we use, but of course we do not know this beforehand. Even if it is better, I'm not proposing that it should replace the algorithm that is currently in use or be added in any official capacity, Instead I would merely like to make it and see how the network performs. Also, I need to finish a project like this for a university course. I could satisfy them with a more boring project, but I'd like to make something potentially useful instead.

After training, the network can implemented without special knowledge by any programmer. The details will depend on the architecture I choose for the network. To give you an idea, one possible architecture would mean two matrix-vector multiplications and applying a certain non-linear function to each component of a vector. I could do the implementing myself if needed, I believe I'm decent programmer. At the moment though I have no experience in modding the game or how the client works, so this task might be easier for someone else.

After two paragraphs we finally come to my point: I need data to train the network. I could probably construct the query for the database myself, but I do not know its structure and I do not have access to it, so I need help from someone.

I quick introduction to neural networks for those who are unfamiliar with the subject:

Spoiler: show

A neural network is trained with a large number of examples that show what kind of response is expected from a given input. After the training, the network is also able to generalize based on the examples it has seen into new situations. Neural networks have successfully been used to get computers do tasks that are traditionally thought to be hard to program, but are relatively easy for humans. Examples include character recognition or classifying websites based on their content.

[Resin_Smoker]

Would be cool to use such knowledge to build an AI that's a worthy opponent in both single and multiplayer matches.

Resin

[ZLO_RD]

That is pretty fun idea... But if you just make 4 separated ratings for every player, back front beach and rock and add the importance of spot you will able to predict outcome without any AI...
Useing an AI for that is just sounds fun anyway...
I also like idea about AI that would just mimic real player gameplay and tatctis... Maybe it is possible to use that neure network for that...

[da_monstr]

Out of curiosity : are you using a custom application (if so, what language are you using) or are you using Matlab?

[zeroAPM]

That is pretty fun idea... But if you just make 4 separated ratings for every player, back front beach and rock and add the importance of spot ...

Front=1
Beach/Rock=5
Back=9.99E99

I suppose this is correct

[ckitching]

In case it's of interest to anyone else, here is the database dump:

https://www.dropbox.com/s/7wtqe7zgkui99 ... sv.gz?dl=0

Go forth and built neural networks and suchlike. Or something else. This represents a few years worth of game history.

[Ionic]

What is the best way to read and sort this data? Where do you see all the players and their spots and where does it list who won?

[ckitching]

Oh, I do beg your pardon, I forgot to post the explanatory accompaniment:

http://sprunge.us/LPcc

The datafile dumped is the output of:

Code: Select all

SELECT
login.login as player,
b.game_type as game_type,
game_player_stats.faction,
game_player_stats.color,
game_player_stats.team,
game_player_stats.place,
game_player_stats.mean,
game_player_stats.deviation,
game_player_stats.after_mean,
game_player_stats.after_deviation,
game_player_stats.score,
game_player_stats.scoreTime,
b.start_time start_time,
b.end_time as end_time,
b.map_name as map_name,
b.map_file as map_file

FROM
    game_player_stats
INNER JOIN
    login
ON
    game_player_stats.playerId = login.id
INNER JOIN
    (
        SELECT
            game_stats.id,
            game_stats.mapId,
            game_stats.startTime start_time,
            game_stats.EndTime as end_time,
            game_stats.gameType as game_type,
            table_map.name as map_name,
            table_map.filename as map_file
        FROM
            game_stats
       INNER JOIN
            table_map
        ON
            table_map.id = game_stats.mapId
    ) as b
ON
    game_player_stats.gameId = b.id


The fields in the output are separated by tab characters. The records are separated by newlines. The first record is the column names.

This format would be suitable for import into prettymuch any database system you might have, and spreadsheet software will probably happily take it as well (if it weren't so enormous, anyway). If you only want a narrow portion of it I suggest you use `grep` or some similar tool to filter it down to a more manageable size before having your way with it.

If you care about game count, notice that that field is implicit in the data set and must be determined by counting records for each player in chronological order. Also note that datestamps are sometimes missing, because ZeP, so your should probably use ID numbers as Lamport clocks if you're going to try and reconstruct that.

[ckitching]

Ah.

Sheeo just informed me I missed a spot.

And by "a spot" I mean "Everything before mid-2014".

There will be a brief interlude while the several GB I missed are extracted

[ckitching]

Here is the complete dump (minus the most recent couple of months):
https://www.dropbox.com/s/mukkqlx15s8ll ... sv.gz?dl=0

Enjoy.