how to train with historic football club data

52 views
Skip to first unread message

Woldo Wondare

unread,
Apr 6, 2020, 6:57:29 AM4/6/20
to Google Research Football
would it be possible to train an agent to replicate a teams performance somehow?
i guess one would need other teams as well, right? to put one team into a context.
what data would you see a good fit for this to work?

Anton Raichuk

unread,
Apr 8, 2020, 2:16:39 PM4/8/20
to Woldo Wondare, Google Research Football
A bit hard to answer questions, since they are too vague.

What team do you want to replicate performance of? And what do you mean by replicating performance at all?

If you want to replicate the performance of a team that is controlled by a trained policy, it is obviously possible since there exists a policy that controls it.
But I don't see a point of doing that.

If you want to replicate performance of built in ai, then maybe, we haven't tried that much, but if the network is expressive enough and you have enough data, i don't see why it would not be possible.

As for data: you can run play_game script enabling dumps, and you'll have plenty of data to train on.

--
You received this message because you are subscribed to the Google Groups "Google Research Football" group.
To unsubscribe from this group and stop receiving emails from it, send an email to google-research-fo...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/google-research-football/719e41f4-3696-4f76-92d6-f03b576432c9%40googlegroups.com.

Woldo Wondare

unread,
Apr 8, 2020, 6:06:39 PM4/8/20
to Google Research Football
i was thinking maybe FC Bayern. If i would take historic data of the german football league, would it be possible to train multiple agents that replicate the strength and weaknesses of those teams and give me a similar expected outcome to their real life strength.
the idea is to simulate this league for cases like this crisis we currently have where there are no real matches happening.
i guess my real question would be: what kind of data would be ideal to represent the performance of real teams? skeletons of players? their individual speeds? or rather more on the team level? (outcomes, percentage of attacks etc.)

you mentioned to generate data by running "play_game" script. is it the one i would control and then dump the data and use this for training?


On Wednesday, 8 April 2020 20:16:39 UTC+2, raveman wrote:
A bit hard to answer questions, since they are too vague.

What team do you want to replicate performance of? And what do you mean by replicating performance at all?

If you want to replicate the performance of a team that is controlled by a trained policy, it is obviously possible since there exists a policy that controls it.
But I don't see a point of doing that.

If you want to replicate performance of built in ai, then maybe, we haven't tried that much, but if the network is expressive enough and you have enough data, i don't see why it would not be possible.

As for data: you can run play_game script enabling dumps, and you'll have plenty of data to train on.

On Mon, Apr 6, 2020 at 12:57 PM Woldo Wondare <locked...@gmail.com> wrote:
would it be possible to train an agent to replicate a teams performance somehow?
i guess one would need other teams as well, right? to put one team into a context.
what data would you see a good fit for this to work?

--
You received this message because you are subscribed to the Google Groups "Google Research Football" group.
To unsubscribe from this group and stop receiving emails from it, send an email to google-research-football+unsub...@googlegroups.com.
Message has been deleted

Anton Raichuk

unread,
Apr 29, 2020, 3:32:46 AM4/29/20
to Woldo Wondare, Google Research Football
Hey Woldo,

The idea you have is very ambitious and sounds like a substantial research project by itself.
Which would not be solved nor even brainstormed over the email.
One possibly relevant research paper is https://arxiv.org/abs/1902.09641 which looks into usage of real sports trajectories and predicting the future.

Another thing I'm confused with, is the relevance of the GRF environment to your idea.
The only connection point I can see is that you want to use GRF as a backend rendering engine to display your trained agents' play.
Which I guess is possible but has nothing to do with research, and likely will still require significant work on the engine side to make physics more real world.

To unsubscribe from this group and stop receiving emails from it, send an email to google-research-fo...@googlegroups.com.

--
You received this message because you are subscribed to the Google Groups "Google Research Football" group.
To unsubscribe from this group and stop receiving emails from it, send an email to google-research-fo...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/google-research-football/b12380e8-dffc-4255-9f91-d7a2d790e7ea%40googlegroups.com.

Woldo Wondare

unread,
Apr 29, 2020, 6:32:41 AM4/29/20
to Google Research Football
looks interesting, thank you.
To unsubscribe from this group and stop receiving emails from it, send an email to google-research-football+unsub...@googlegroups.com.

--
You received this message because you are subscribed to the Google Groups "Google Research Football" group.
To unsubscribe from this group and stop receiving emails from it, send an email to google-research-football+unsub...@googlegroups.com.
Reply all
Reply to author
Forward
0 new messages