Specify a player to build a team around?

[mildfuzz]

My specific use case is that I believe adding Ronaldo next game week is the best play, and I believe that, from how I understand the algorithm to work, the system is incapable of making that call because there is no premier league data supporting that.

Can we then give the system a nudge and insist on certain players in certain weeks?

Or...am I mistaken? and we are importing data from foreign leagues and the system has just decided it's best to leave him out? Do we have a process for assessing foreign imports?

[nbarlowATI]

Hi @mildfuzz , you're not mistaken - the system as it is now doesn't know anything about players like Ronaldo... (Just saw he's been priced at 12.5M - wow!!)

There are actually a couple of separate-but-related issues:

• Firstly the PlayerModel, which gives probabilities for a player to score, assist, or do neither, for every goal his team scores, doesn't have past data for Ronaldo, so would just assign him the average probabilities for all FWDs. It would be interesting to see if we could find a data source to import the data we'd need from other leagues - e.g. presumably there is also a Fantasy Serie A - I wonder if they have an accessible API.... (of course even if we did manage to build up a pan-European dataset of players, it wouldn't take into account the "wet Tuesday night in Stoke" question! :) )
• Secondly, the prediction of how many minutes a player will be on the pitch is currently very simplistic, and just looks back at how many minutes they played in the last few PL games. This prediction in general is probably where there is most room for improvement in AIrsenal - not just for new players (re)joining the league, but also for periods where there is fixture congestion, and/or important cup games - how many minutes a player played in the last 3 matches is not always the best predictor of whether they'll start the next one. Any ideas for improving this would be very welcome!
• Coming back to the original question - could we introduce a way of saying we definitely want a certain player, even if the algorithm doesn't predict (m)any points for them? I guess it shouldn't be too hard - a new command-line option in the optimization to pass through a list of player_ids, and then have the "make_best_transfers" and related functions aware of this.. It could be a bit fiddly, but definitely not impossible... We can add this to a list of features that we'll look at implementing when we have time (unless you want to have a go at it in the meantime? :) )

[mildfuzz]

If I could find time, would certainly have a bash. Python not my strongest suit though.

Could another way to manage this be to somehow overestimate new signings?

[jack89roberts]

I can imagine a few different versions of this:

• I want to transfer in player(s) X, Y, Z (e.g. Ronaldo), who should I transfer out? (I think this is what you're currently hoping for?)
• I want to transfer out player(s) X, Y, Z. Who should I transfer in?
• I want to create a new (wildcard, free hit...) squad that has player(s) X, Y, Z in it, and doesn't have player(s) A, B, C in it.
• I do/don't want player(s) X, Y, Z in my squad in gameweek N.

These would all be interesting for a human-in-the-loop version of AIrsenal and would be fun to play with I think. Like Nick says they should all be possible, but probably would take a decent amount of time/thought to incorporate them and get it working.

The issue with AIrsenal not having good predictions for new players would need more data and an improved model to fix. Definitely on the list of things we'd like to do though and contributions/ideas welcome!

[Tdarnell]

I can imagine a few different versions of this:

• I want to transfer in player(s) X, Y, Z (e.g. Ronaldo), who should I transfer out? (I think this is what you're currently hoping for?)
• I want to transfer out player(s) X, Y, Z. Who should I transfer in?
• I want to create a new (wildcard, free hit...) squad that has player(s) X, Y, Z in it, and doesn't have player(s) A, B, C in it.
• I do/don't want player(s) X, Y, Z in my squad in gameweek N.

These would all be interesting for a human-in-the-loop version of AIrsenal and would be fun to play with I think. Like Nick says they should all be possible, but probably would take a decent amount of time/thought to incorporate them and get it working.

The issue with AIrsenal not having good predictions for new players would need more data and an improved model to fix. Definitely on the list of things we'd like to do though and contributions/ideas welcome!

How about using the understat scraper to acquire data for all different leagues? This can also give an idea for not only players xG but also their typical position on the pitch: https://github.com/douglasbc/scraping-understat-dataset