# Dominion Strategy Forum

• August 14, 2020, 10:50:16 pm
• Welcome, Guest

### News:

DominionStrategy Wiki

Pages: [1]

### AuthorTopic: Cute idea? Piggy back on qvist's rankings for pop buys/supply win  (Read 1627 times)

0 Members and 1 Guest are viewing this topic.

#### rrenaud

• Offline
• Posts: 990
• Uncivilized Barbarian of Statistics
• Respect: +1184
##### Cute idea? Piggy back on qvist's rankings for pop buys/supply win
« on: January 25, 2012, 02:08:27 pm »
0

Especially once Qvist is done with his report on the community rankings, there is a nice golden dataset for card rankings.

Using that data as the desired outcome, I (we, you?) could do some regression to find weights from features like percentage of games a card was purchased, win rate given any purchased, win rate per purchase, average gains per game.  This would determine how important each feature is to the rankings.  Then using those weights, for example, we could compute personalized rankings (rankings given a single players stats) and card synergies (change in score given a new card is available).
Logged

#### Qvist

• Mountebank
• Offline
• Posts: 2400
• Respect: +4076
##### Re: Cute idea? Piggy back on qvist's rankings for pop buys/supply win
« Reply #1 on: January 26, 2012, 03:36:35 am »
0

Sure it would be cute as most of us here love statistics. But I doubt a little bit that you really get a significant specific formula (specific coefficients) out of these parameters to evaluate that. But you can prove me the opposite if you want

#### rrenaud

• Offline
• Posts: 990
• Uncivilized Barbarian of Statistics
• Respect: +1184
##### Re: Cute idea? Piggy back on qvist's rankings for pop buys/supply win
« Reply #2 on: January 26, 2012, 08:04:36 pm »
0

I mean, I certainly won't be able to reproduce the ordering exactly with 4 coefficients, but it  will probably produce something reasonable.
Logged

#### rrenaud

• Offline
• Posts: 990
• Uncivilized Barbarian of Statistics
• Respect: +1184
##### Re: Cute idea? Piggy back on qvist's rankings for pop buys/supply win
« Reply #3 on: February 13, 2012, 03:58:05 am »
+2

I played with this a bit.  The code is here: https://github.com/rrenaud/dominionstats/tree/master/card_ranker

I've just been using the scipy generic optimization function that doesn't use any gradient information, and it has seemed to do okay.  I've just been using it to try to minimize pairwise inversions against Qvists list.  But doing something smarter with a proxy loss on the accuracy so that I can get a smooth derivative might be worthwhile.

If you have a statsy bend, I can make this painless to play around with for R or something.  It's really just a small couple data files (one for the card stats, one for the rankings).

With just one feature (win rate given any avail), I get 86% of the pairings correct.  This is kind of suspicious, and sort of leads me to believe the serpent is eating its tail, that is, QVist's rankings might be pretty biased towards CR's ratings itself.

With some small weight on the frequency a card is gained, I get it to 87%.
Adding is_vp, num_plus_actions, and is_reaction gets the accuracy to 90%, but is perhaps dangerously overfit.

The currently induced rankings:

2 Secret Chamber,Herbalist,Duchess,Native Village,Pearl Diver,Cellar,
Fool's Gold,Lighthouse,Chapel

3 Chancellor,Workshop,Smugglers,Woodcutter,Develop,Fortune Teller,Oracle,Black Market,
Scheme,Silver,Tunnel,Loan,Swindler,Steward,Warehouse,Fishing Village,

4 Thief,Coppersmith,Bureaucrat,Pirate Ship,Talisman,
Conspirator,Spice Merchant,Militia,Monument,Remake,
Jack of All Trades,Salvager,Bishop,Caravan,Young Witch,Sea Hag,Tournament

5 Counting House,Saboteur,Contraband,Cache,Duke,Mandarin,Outpost,Stash,Explorer,
Tribute,Harvest,Mine,Horn of Plenty,Royal Seal,Inn,Library,Council Room,Merchant Ship,