Win Time Study - Pro Stats for Pro Grinders.

[Scaredgirl]

We know the probability of winning a card in a spin. We know the value of cards in AI3 decks. We can easily calculate the average revenue generated by spins.

Well, this seems like a good way to prove your argument, SG.

Yeah, I'd be interested to in these statistics.

Honestly the only way to do this right is to do 3-5 hour runs with and without spins with a fast EM deck and with the fastest deck, say V2 fire rush and compare notes.  That is an entirely different study though.

There's no need to do any 3-5 hour runs. The whole point is that you can do it without any actual playing, simply by using math.

All we need to do is calculate the average revenue generated by spins. We already have all the different parts of the equation, rest is simple math. After we get average revenue from 3 spins (for example 17 electrum) we simply add that number on top of each win.

Yeppers, doing that as we speak. I've got 230 games logged with specific data for spins and I'm going to go to 500. Hopefully by tomorrow, I'll have all stats needed from the 500 and post the info on the forums. Then we can use more specific numbers when doing the math for all the data from this particular study

You are actually doing the opposite of what I was talking about.

Problem with your method is that the sample size is way too small, and margin of error is way too high, to give any statistically relevant results. Although 500 sounds like a big number, in statistics it's peanuts. You would probably need tens of thousands of spins to make it legit.

What I suggested was that we can calculated the average revenue without playing a single game, simply using the data we have about card cost and win probability. I'm sure there is at least one math nerd here who is willing to do it.

Think of it like this:

We have a game where you throw a single dice. We want to determine how often we throw a 4, 5 0r 6. Sure we could throw a 1000 times and count how many times we get a 4, 5 or 6, but not only would it take a long time, it would also give inaccurate results. It's much better to use math and find out it's a 50% chance. This way there is zero margin of error and everything is more "scientific".

[Kuross]

We know the probability of winning a card in a spin. We know the value of cards in AI3 decks. We can easily calculate the average revenue generated by spins.

Well, this seems like a good way to prove your argument, SG.

Yeah, I'd be interested to in these statistics.

Honestly the only way to do this right is to do 3-5 hour runs with and without spins with a fast EM deck and with the fastest deck, say V2 fire rush and compare notes.  That is an entirely different study though.

There's no need to do any 3-5 hour runs. The whole point is that you can do it without any actual playing, simply by using math.

All we need to do is calculate the average revenue generated by spins. We already have all the different parts of the equation, rest is simple math. After we get average revenue from 3 spins (for example 17 electrum) we simply add that number on top of each win.

Yeppers, doing that as we speak. I've got 230 games logged with specific data for spins and I'm going to go to 500. Hopefully by tomorrow, I'll have all stats needed from the 500 and post the info on the forums. Then we can use more specific numbers when doing the math for all the data from this particular study

You are actually doing the opposite of what I was talking about.

Problem with your method is that the sample size is way too small, and margin of error is way too high, to give any statistically relevant results. Although 500 sounds like a big number, in statistics it's peanuts. You would probably need tens of thousands of spins to make it legit.

What I suggested was that we can calculated the average revenue without playing a single game, simply using the data we have about card cost and win probability. I'm sure there is at least one math nerd here who is willing to do it.

Think of it like this:

We have a game where you throw a single dice. We want to determine how often we throw a 4, 5 0r 6. Sure we could throw a 1000 times and count how many times we get a 4, 5 or 6, but not only would it take a long time, it would also give inaccurate results. It's much better to use math and find out it's a 50% chance. This way there is zero margin of error and everything is more "scientific".

Don't get me wrong, I appreciate the mathematical aspect of a good statistical approach, but even in statistics, there is a +/- variance to any number generated within a statistical model. I figure with 1500 spins, my numbers will fall within that variance and be as close to valid as one can get in statistics.

I honestly believe, with myself having over 6000 games of AI3 under my belt, and this particular sample, the numbers are a good representation of what one can expect over a good length of time playing AI3.

[tinkady]

can we just get a list of the spin procedures from zanz? and figure it out mathematically like scaredgirl said. maybe its just a matter of looking at all the AI3 decklists and calculating chances of getting each card?
I think it could take a while to get a good average...with the 500 spins youve done did it approach a certain limit?
i'm a bit of a math nerd, I could probably do it if somebody gave me all the necessary information...unless its just a buttassload of information in which case i'm too lazy 

[wavedash]

Don't get me wrong, I appreciate the mathematical aspect of a good statistical approach, but even in statistics, there is a +/- variance to any number generated within a statistical model. I figure with 1500 spins, my numbers will fall within that variance and be as close to valid as one can get in statistics.

I honestly believe, with myself having over 6000 games of AI3 under my belt, and this particular sample, the numbers are a good representation of what one can expect over a good length of time playing AI3.

I'm with Kuross here. Even if there's some error (and there probably is), it's small enough that it won't matter.

Besides, it's good to have a practical answer to compare to the theoretical.

[tinkady]

but we don't even have a theoretical answer yet...and if we can get a theoretical answer its probably more accurate than an experimental

[Kuross]

but we don't even have a theoretical answer yet...and if we can get a theoretical answer its probably more accurate than an experimental

I'm assuming "experiemntal" = actual data collected, given the argument is in favor of a theorectical model.

If any theoretical model has an average dramatically over/under 24.25 electrum per game in spins, then either I was extremely unlucky/lucky in spins over 500 games (1500 spins), or the model is off. However, theoretical models should be able to explain, or predict, actual data to some accuracy. Since the goal of any statistical approach is to accurately predict a given variable then that data should reflect actual data recorded. However, if the statistics don't closely reflect data in a same, or similar, sample, then it's the theoretical model that is off since actual data trumps theortical.

So, if any model were to be created to account for electrum gained in spins for Elder, that model should be accurate enough to account for my data collected, with a certain amount of +/- variances, with those variances within acceptable standards.

Long story short, 24.25 should be a good enough approximation of electrum won from spins over time.

[jmdt]

but we don't even have a theoretical answer yet...and if we can get a theoretical answer its probably more accurate than an experimental

I'm assuming "experiemntal" = actual data collected, given the argument is in favor of a theorectical model.

If any theoretical model has an average dramatically over/under 24.25 electrum per game in spins, then either I was extremely unlucky/lucky in spins over 500 games (1500 spins), or the model is off. However, theoretical models should be able to explain, or predict, actual data to some accuracy. Since the goal of any statistical approach is to accurately predict a given variable then that data should reflect actual data recorded. However, if the statistics don't closely reflect data in a same, or similar, sample, then it's the theoretical model that is off since actual data trumps theortical.

So, if any model were to be created to account for electrum gained in spins for Elder, that model should be accurate enough to account for my data collected, with a certain amount of +/- variances, with those variances within acceptable standards.

Long story short, 24.25 should be a good enough approximation of electrum won from spins over time.

Good enough for me.

[wavedash]

To add to this, a sample size of 1200 has a margin of error or +/-3%.

[jmdt]

The first unupped decks are added. More to come.

[jmdt]

I have time for some more testing this week.  Will post more unupped decks and the data from my analysis spreadsheet.

[hrmmm]

i was a bit tired of testing, but finally...
more stats:
Mono-Death Mummy Rush (http://elementscommunity.org/forum/index.php/topic,10503.0.html)
Games:   50
   
avg. Sec/Game:   81.100
Sec/Turn:   9.541
Clicks/Game:   13.920
Clicks/Turn:   1.638
Score/Min:   9.307
Electrum/Min:   9.307
   
Time overall (min.):   67.583
Wins:   47
Ems:   0
Losses:   3
Score:   629
Electrum:   629
   
   
Avg ttw:   8.500
   
ttw Details:   

ttw 6:   1
ttw 7:   11
ttw 8:   13
ttw 9:   15
ttw 10:   7
ttw 11:   3

[Scaredgirl]

As an academic person I have to try to find flaws in this study to make it better.

If you look at the results, you can see that decks tested by jmdt have the most top spots, while decks tested by hrmmm and kuros have the bottom spots. 5 decks in top-6 were tested by jmdt, while only 1 deck in bottom-7 was tested by jmdt. There looks to be clear correlation between time to win and whether or not it was tested by jmdt.

So my question is, do decks tested by jmdt get better score because of computer performance or style of playing? Or did jmdt just happen to pick all the fast decks by chance, while others happened to pick slow decks?

Wouldn't this study be more "scientific" if only one person tested everything? (I'm not saying a one person should do it because it's a huge task).