search

Skybox-Technologies / csgo-ai-competition

CSGO.ai competition
34 stars 16 forks source link

Model Prediction #9

Open jasonjiang09 opened 4 years ago

jasonjiang09 commented 4 years ago

Hello,

I was just wondering if we are betting our model to predict the match or predict the round. Also, up to which point of the round/match is the data no longer allowed for the prediction? For instance, is it up to the end of the freeze period of the round that we get to use the info to predict the round?

In addition, will we later be receiving data on the time of the match, which team is playing, and which players are playing in the dataset like should we consider player's performances and what teams are playing in the model? Also, will we know which patch the dataset is from?

Thanks,

Jason

kristoff3r commented 4 years ago

Hi,

I'm not sure I understand your first question. The idea is that, at any time during a round, we provide your model with a snapshot which it uses to predict who will win the round. Of course accuracy will be lower in freezetime than just before the bomb explodes, but that's just how it is.

I decided against providing ways to identify teams/players for practical reasons. While it would be interesting and probably increase accuracy for professional games, it would also make the models less useful for new and lower tier players.

I don't have the time of the match, but I could perhaps add patch numbers to the samples if you can convince me that it is useful. The games in the current dataset are on different but recent patches, and no games are older than 2019.

jasonjiang09 commented 4 years ago

Ok got it thanks for the feedback!

So I think the patch numbers are important because for instance during the beginning of 2019, the Aug weapon was considerably better than it is now. The patterns popped up for other weapons and also nerfing weapons as well. In addition, the economy shifted significantly based on the patch. For instance, the terrorist has a better buy now than at some point in the past.

Also, to clarify the question. I also meant if like say we're predicting 5th round at 30 seconds into the game, will we be able to use info from the previous rounds to determine the outcome of the 5th round.

andrew-son commented 4 years ago

@kristoff3r Don't know if it convinces you but I would find patch numbers useful at least for one case -- out of all maps de_vertigo underwent a lot of changes to its' structure. For now I compare default spawns with the player positions at FreezeTime entries and then write the patch version to the rows after (since I assume events are in chronological order). A bit gimmicky and not sure if it's working 100% of the times. Patch version would make this much simpler!

Speaking of general purposes: I think it's less about economy patches (since current economy patch was on March 13, 2019 with the tweak on May 14, 2019 (loss number can't be more than 5) which is fully covered as you said there are 2019-2020 demos. Mostly, about weapon patches (like aforementioned aug/krieg meta) and map updates.