Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[ehunt]

If you click on the "card winningness" on councilroom, you'll notice that menagerie is a good card. OK, definitely.

Then you'll notice that among all the <5 cost villages, the single best fit for menagerie is worker's village. Not hamlet, not fishing village. (This comparison show up in the default that rrenaud put in there if you click that link on councilroom - you don't have to ask it anything.) I have no explanation! I guess +buy is nice, but hamlet gives that and also causes you to draw a billion cards (admittedly after possibly making your hand slightly worse).

[DG]

The extra buys suits engines. The villages let you play terminals out of hand with an action remaining for the menagerie. The hamlet presumably has to throw cards away to be useful, eating up the cards you draw with the menagerie.

[rrenaud]

You mean this?

http://councilroom.com/supply_win?&targets=Menagerie&interaction=Actions%3E%3D2%26%26Cost%3C%3D4&nested=false&unconditional=true

The card stats might tend to over-rate cards with +buys.  Since being purchased looks good according to the stats, cards which help purchasing/gaining look good.  Of course, the win given any gain stats are good for menagerie beyond just the %+, so maybe it's legit.

I'd tend to like Fishing Village and Hamlet more as Menagerie supports than Worker's Village though.

[ehunt]

The card stats might tend to over-rate cards with +buys.  Since being purchased looks good according to the stats, cards which help purchasing/gaining look good.  Of course, the win given any gain stats are good for menagerie beyond just the %+, so maybe it's legit.

If I understand you correctly, the +buy in worker's village inflates Delta qual because part of what qual is measuring is how much more likely you are to gain stuff? Shouldn't hamlet blow worker's village out of the water, just because they tend to cause you to gain like infinity of them? This is a super-cool app by the way, +1.

[popsofctown]

I think since you need to discard TWO cards to get Hamlet to do everything Worker's Village does, the Worker's Village does well. 


Also, remember, correlation doesn't imply causation.  If Hamlet is on the board with Menagerie, you're probably going to try to add Hamlet to your Menagerie deck, even if you don't have an outlet for +actions.  If you are playing a Menagerie deck and Worker's Village is on the board, you're only going to buy the Worker's Village if you have a an outlet for your actions. 

So the Hamlet Menagerie sample is going to be nearly every game they appear together, the WV Menagerie sample is going to be a pruned set of games where great terminals are available to power the Menagerie engine to maximum wowsome.

[rrenaud]

The conditional ranks for menagerie are based on the menagerie stats when other cards are available.  Whether or not the other cards are purchased not directly relevant to the delta for menagerie.  Of course, you can do the converse, and see how much menagerie helps the other cards, and that won't have this bias.

For the existing ranking function is primarily based on win rate given avail and chance of being gained.  I think those two alone got 86% of the pairs ranked correctly. 

Code: [Select]

  // This function was found by card_ranker/optimize_ranks.py
  return (52.926 * win_rate_given_gain +
          1.358 * log_odds_any_gained +
          -1.161 * num_plus_actions +
          -1.712 * is_vp +
          1.625 * is_reaction);

I had some compensating factors for things like +actions (that tend to be gained more often) and reactions (that tend to be gained when your opponent is playing well, eg, attacking) and is_vp (winners tend to get lots of vp even if the vp itself is the effect rather than the cause of the dominant position).  With the three mitigating factors, the accuracy was at about 90%.

But they were optimized against QVist's single card ranks.  I could definitely see the need to add some additional mitigating factors for the combo ranking.  Like here, cards that make it easy (+buy/gain) or desirable to have many cards in the deck (fairgrounds) might be looked too highly upon by the current conditional ranking function.