Print

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

[yuma]

I had a thought about using Councilroom data for game predictions and wonder if it has any useful application, as I write this I am not sure if it does but am simply putting the thought out there.

For example, if I wanted to predict who would win the last annotated game on theory's blog between him and Captain Frisk could I use councilroom data?  Here is what I mean:

In that game the following cards were used and here are theory's and Captain_Frisk's councilroom stats for each card, the stat I am using is the "win rate given availability" stat:

                        theory              captain_frisk
Scrying Pool       1.25 (.20)           1.21 (.12)
Scheme             1.06 (.33)           1.29 (.14)
Warehouse         1.26 (.11)           1.37 (.09)
Apprentice          1.08 (.25)           1.33 (.10)
Harvest             1.15  (.20)          1.29 (.12)
Ambassador       1.29 (.14)           1.36 (.09)
Tunnel               .91 (.42)            1.21 (.15)
Conspirator        1.11 (.12)           1.33 (.09)
Counting House   1.17 (.12)           1.29 (.09)
Bank                 1.22 (.12)           1.32 (.09)

average             1.15                  1.30

As a result I would pick Captain_Frisk to win the game, which he did 55-36.

Here is another example with me against tlloyd in the most recent IsoDom tournament:

                             tlloyd                yuma
Baron                 1.28 ± 0.12         1.30 ± 0.16
Conspirator         1.10 ± 0.13         1.42 ± 0.15
Expand              1.17 ± 0.12          1.35 ± 0.17
Fortune Teller     1.15 ± 0.14          1.08 ± 0.16
Goons                1.22 ± 0.12         1.17 ± 0.20
Grand Market      1.15 ± 0.12         1.12 ± 0.18
Mountebank        1.14 ± 0.13         1.11 ± 0.19
Quarry               1.11 ± 0.13         1.18 ± 0.16
Royal Seal         1.15 ± 0.13          1.23 ± 0.17
Walled Village     1.10 ± 0.15         1.25 ± 0.17

 average           1.157                   1.221

So if I used this as a predictor I would suggest I would win, but I lost this game 59-44, so obviously it isn't a perfect predictor, but no predictor is.

I guess what I am asking is since stats isn't my best subject in the world does this method have any application? If so is there anyway to make it a less time consuming venture, putting these two games took me about 10 minutes in copying and pasting and such (I really need to get a mouse for my laptop).

If there is no useful application is there anyway to make something useful out of it? If no to both of those questions then feel free to ignore.

[rod-]

It would be marginally interesting to go back over tournament results and use this predictor as a comparison against the baseline (level) predictor for the results, and it would be relatively simple to script it out.

However, until the next tournament comes out and people start betting on their tournament brackets, it's of no practical value.  I'd even argue that its' academic value is tainted by virtue of the fact that the "training set" of data (all games by player X on CR) is included within the "testing set" of data (some specific games by player X, also on CR)

[rrenaud]

The way I see it, this is just a more nuanced version of trueskill that gives players a per card skill that is combined to get a kingdom specific game skill.

You can do a "online" evaluation of the ranking method, where you give the system the previous games, and then ask it to predict the next one.  Or you can use a hold out set.  There are plenty of fair ways to see if such a method is more predictive than a different baseline predictor.

You can even look at this as another way of mitigating card selection biases, since you can form an aggregate skill measure that was trained on a biased distribution (see Obi Wan) by computing the per card skills with fair weights.

It could be interesting, and I'd happily help someone who wanted to try it.  But I'd rather try to do other things among the countless cool analysis ideas for cr.

[Obi Wan Bonogi]

Comparing winrates from player to player will only take you so far as they don't take into account disparity in opponent level during that history.  I would suggest that a more accurate analysis along the lines you are looking for would be to: compare the player's winrate for individual cards in the set to their overall winrate and then do the same for the opponent. Then compare those deviations to get your expectation.