|
On March 21 2008 10:32 IzzyCraft wrote:
no one really cares about the zvz match up its basically you overpool or 9 pool or your pretty much dead after taht its just your choice of agression fest lings, muta, sourge, devoirers. When i play zvz i think very little of my build after the 9 pool i just do what come naturally which is fast lair to zerg air when i play
Just because you only use two builds doesn't mean that only two builds work.
|
On March 21 2008 09:19 5HITCOMBO wrote:
Guys, how often are zvz games at the pro level decided off of the build?
It's honestly more about adaptation. Every zvz is different.
Prob half end mere minutes after the openings have been chosen.
I dont know if I think adaptation is a great description. More like tactical exploitation. Given a game is on even or semi-even footing build-wise whomever can "read" their opponent better in a fairly arbitrary/guessing manner then execute based on that read is the winner. And im not talking about build here.
If you are making nonstop ling in a 12hat vs 12hat opening situation and your opponent makes a few drones at the start of the ling pump cycle, you dont know he has made drones until they actually hatch. Sure after you see them you could decide to react in response but given that you are both producing your larvae, tech and everything at the exact same timing it is logical that the match up is related mostly to guesswork.
Taking the initiative is also often disastrous being that the person defending should in the army vs army sense(Not taking into account base harassment) always have the advantage.
|
On March 21 2008 23:34 red.venom wrote:Show nested quote +On March 21 2008 09:19 5HITCOMBO wrote:
Guys, how often are zvz games at the pro level decided off of the build?
It's honestly more about adaptation. Every zvz is different.
Prob half end mere minutes after the openings have been chosen.I dont know if I think adaptation is a great description. More like tactical exploitation. Given a game is on even or semi-even footing build-wise whomever can "read" their opponent better in a fairly arbitrary/guessing manner then execute based on that read is the winner. And im not talking about build here. If you are making nonstop ling in a 12hat vs 12hat opening situation and your opponent makes a few drones at the start of the ling pump cycle, you dont know he has made drones until they actually hatch. Sure after you see them you could decide to react in response but given that you are both producing your larvae, tech and everything at the exact same timing it is logical that the match up is related mostly to guesswork. Taking the initiative is also often disastrous being that the person defending should in the army vs army sense(Not taking into account base harassment) always have the advantage. Jaedong..if half are decided once the builds are chosen, why does he win so much? Why are there people who are 'good' at zvz
|
Russian Federation4235 Posts
Probabilities and StarCraft don't mix.
Or, rewording, playing unexpected and playing random are two completely different things, of which only one is productive.
|
I have not explained the concept clearly. I'll try again by using a simpler example.
In rock paper scissors, rock > scissors > paper > rock. In the probability table, it would be listed as
r p s
r 50 0 100
p 100 50 0
s 0 100 50
Now you say rock v rock is not 50% win, it is a tie, well you rego and in that rego you have a 50% shot. My program would say you should pick each randomly with 33% chance. This means that this is the only strategy such that there exist no other strategy that beats it more than 50% of the time. This is a trivial example, so lets say rock paper scissors is a little more complicated, lets say if one player chooses rock and the other scissors, this is a "skunk". A skunk is like 3 wins, so if you were doing best of 10 this is much better than a normal win.
What percent of the time should you choose each now? Try to answer this now in your head (answer below). I guarantee no one will even come close to the correct answer (except maybe some poker players) and that is the point of this topic, to try and calculate something that we all do naturally but much more accurately.
Alot of people have said "but starcraft is not about the build its about how you play." This is totally true, but the fact remains some builds do have small advantages over the other. Even if these advantages are small, one should still strive to cache in on every advantage possible no matter how small, that is starcraft.
Starcraft is just like rock paper scissors, except by choosing a counter build you don't automatically win, instead you might have a 55% chance of winning. Where as some other counter like 9pool versus 12 hatch might have an 85% chance of win. Just like the "skunk" from rock paper scissors, choosing a 9pool over a 12hatch has a much higher pay off than something like 12 pool over a 9 pool.
Obviously skill has a much higher impact than the build in many cases, if you are really serious about improving you should be practicing not reading this. But this topic is theoretical for players already at "the pro level". I believe that even though this isn't directly that useful, if given reasonably accurate estimation of the win percentages, it will shed some light on which zerg builds are viable and which are not, perhaps better than previous analysis has done.
Here is the answer to the rock paper scissors skunk thing:
The answer is 20% rock, 60% paper, and 20% scissors. You are probably thinking pssh that can't be right, but it is actually easy to verify that it is correct using the assertion mentioned earlier: "this is the only strategy such that there exist no other strategy that beats it more than 50% of the time." Although since weight of winning with rock over scissors is +3 instead of 1, it needs to be reworded to "this is the only strategy such that there exist no other strategy that has a positive expected score against it."
Here's how to check it.
rock v the strat:
strat choose rock 20%, but score diff is 0 since tie
strat chooses paper 60%, so 0.6 * -1 (for loss)
strat chooses scissors 20%, so 0.2 * 3 (3 for skunk)
add them up ======================================
equals 0
This means that if they choose rock against your strat on average the net score difference is 0.
Repeat for paper: 20% * 1 + 60% * 0 + 20% * -1 = 0
Repeat for scissors: 20% * -3 + 60% * 1 + 20% * 0 = 0
This means that no matter what they choose they cannot beat you on average. You are probably thinking that's nice, but even if they have an inferior strategy, using this will not actually give you an advantage, it will just tie. That is exactly correct. This only tells you how to be unbeatable on average not how to exploit other players tendencies. If you know your opponents tendencies, you can calculate which build you should use against them by just choosing the one that has the highest expected win average. The only time it pays to use a randomized starting build, is when your opponent knows your tendencies and you want to give him no option that on average beats you.
For all this to work we need better estimations of one build versus another. This is where I need your help. You don't have to redo the whole table, if you just say stuff like "I think at the pro level 9 pool has a 40% chance of beating 12 pool." If you are someone awesome like incontrol or midian you don't even need to explain why (other people do though).
|
On March 22 2008 02:14 flag wrote:
I have not explained the concept clearly. I'll try again by using a simpler example.
In rock paper scissors, rock > scissors > paper > rock. In the probability table, it would be listed as
r p s
r 50 0 100
p 100 50 0
s 0 100 50
Now you say rock v rock is not 50% win, it is a tie, well you rego and in that rego you have a 50% shot. My program would say you should pick each randomly with 33% chance. This means that this is the only strategy such that there exist no other strategy that beats it more than 50% of the time. This is a trivial example, so lets say rock paper scissors is a little more complicated, lets say if one player chooses rock and the other scissors, this is a "skunk". A skunk is like 3 wins, so if you were doing best of 10 this is much better than a normal win.
What percent of the time should you choose each now? Try to answer this now in your head (answer below). I guarantee no one will even come close to the correct answer (except maybe some poker players) and that is the point of this topic, to try and calculate something that we all do naturally but much more accurately.
Alot of people have said "but starcraft is not about the build its about how you play." This is totally true, but the fact remains some builds do have small advantages over the other. Even if these advantages are small, one should still strive to cache in on every advantage possible no matter how small, that is starcraft.
Starcraft is just like rock paper scissors, except by choosing a counter build you don't automatically win, instead you might have a 55% chance of winning. Where as some other counter like 9pool versus 12 hatch might have an 85% chance of win. Just like the "skunk" from rock paper scissors, choosing a 9pool over a 12hatch has a much higher pay off than something like 12 pool over a 9 pool.
Obviously skill has a much higher impact than the build in many cases, if you are really serious about improving you should be practicing not reading this. But this topic is theoretical for players already at "the pro level". I believe that even though this isn't directly that useful, if given reasonably accurate estimation of the win percentages, it will shed some light on which zerg builds are viable and which are not, perhaps better than previous analysis has done.
Here is the answer to the rock paper scissors skunk thing:
The answer is 20% rock, 60% paper, and 20% scissors. You are probably thinking pssh that can't be right, but it is actually easy to verify that it is correct using the assertion mentioned earlier: "this is the only strategy such that there exist no other strategy that beats it more than 50% of the time." Although since weight of winning with rock over scissors is +3 instead of 1, it needs to be reworded to "this is the only strategy such that there exist no other strategy that has a positive expected score against it."
Here's how to check it.
rock v the strat:
strat choose rock 20%, but score diff is 0 since tie
strat chooses paper 60%, so 0.6 * -1 (for loss)
strat chooses scissors 20%, so 0.2 * 3 (3 for skunk)
add them up ======================================
equals 0
This means that if they choose rock against your strat on average the net score difference is 0.
Repeat for paper: 20% * 1 + 60% * 0 + 20% * -1 = 0
Repeat for scissors: 20% * -3 + 60% * 1 + 20% * 0 = 0
This means that no matter what they choose they cannot beat you on average. You are probably thinking that's nice, but even if they have an inferior strategy, using this will not actually give you an advantage, it will just tie. That is exactly correct. This only tells you how to be unbeatable on average not how to exploit other players tendencies. If you know your opponents tendencies, you can calculate which build you should use against them by just choosing the one that has the highest expected win average. The only time it pays to use a randomized starting build, is when your opponent knows your tendencies and you want to give him no option that on average beats you.
For all this to work we need better estimations of one build versus another. This is where I need your help. You don't have to redo the whole table, if you just say stuff like "I think at the pro level 9 pool has a 40% chance of beating 12 pool." If you are someone awesome like incontrol or midian you don't even need to explain why (other people do though).
I do not understand how you can choose a strategy for a game of equal random chance.
Are you saying that you understand how to win every rock paper scissors match on average?
|
No, im saying I have a strategy for rock paper scissors that no one can beat me more than half the time. That strategy is choose rock 1/3 the time, choose scissors 1/3 of the time, choose paper 1/3 of the time, make this choice randomly.
I'm also saying that this is the only strategy that has this property.
|
I also disagree with which BOs you include. Nobody ever 8pools, since it makes no sense with respect to overlord timing. Also, the build I use the majority of the time is 11gas 10pool.
|
On March 22 2008 04:18 azndsh wrote:
I also disagree with which BOs you include. Nobody ever 8pools, since it makes no sense with respect to overlord timing. Also, the build I use the majority of the time is 11gas 10pool.
That means you agree actually, since after computation it was deemed don't 8 pool. It doesn't effect the percentages if the build isn't used, so it can't hurt to have them in extra. Please state your estimation analysis of 11gas 10pool versus other builds like 12pool 9pool 12hatch. And I'll throw it in in the chart, I've seen other zerg do this alot too so it should be added.
|
On March 22 2008 03:25 5HITCOMBO wrote:Show nested quote +On March 22 2008 02:14 flag wrote:
I have not explained the concept clearly. I'll try again by using a simpler example.
In rock paper scissors, rock > scissors > paper > rock. In the probability table, it would be listed as
r p s
r 50 0 100
p 100 50 0
s 0 100 50
Now you say rock v rock is not 50% win, it is a tie, well you rego and in that rego you have a 50% shot. My program would say you should pick each randomly with 33% chance. This means that this is the only strategy such that there exist no other strategy that beats it more than 50% of the time. This is a trivial example, so lets say rock paper scissors is a little more complicated, lets say if one player chooses rock and the other scissors, this is a "skunk". A skunk is like 3 wins, so if you were doing best of 10 this is much better than a normal win.
What percent of the time should you choose each now? Try to answer this now in your head (answer below). I guarantee no one will even come close to the correct answer (except maybe some poker players) and that is the point of this topic, to try and calculate something that we all do naturally but much more accurately.
Alot of people have said "but starcraft is not about the build its about how you play." This is totally true, but the fact remains some builds do have small advantages over the other. Even if these advantages are small, one should still strive to cache in on every advantage possible no matter how small, that is starcraft.
Starcraft is just like rock paper scissors, except by choosing a counter build you don't automatically win, instead you might have a 55% chance of winning. Where as some other counter like 9pool versus 12 hatch might have an 85% chance of win. Just like the "skunk" from rock paper scissors, choosing a 9pool over a 12hatch has a much higher pay off than something like 12 pool over a 9 pool.
Obviously skill has a much higher impact than the build in many cases, if you are really serious about improving you should be practicing not reading this. But this topic is theoretical for players already at "the pro level". I believe that even though this isn't directly that useful, if given reasonably accurate estimation of the win percentages, it will shed some light on which zerg builds are viable and which are not, perhaps better than previous analysis has done.
Here is the answer to the rock paper scissors skunk thing:
The answer is 20% rock, 60% paper, and 20% scissors. You are probably thinking pssh that can't be right, but it is actually easy to verify that it is correct using the assertion mentioned earlier: "this is the only strategy such that there exist no other strategy that beats it more than 50% of the time." Although since weight of winning with rock over scissors is +3 instead of 1, it needs to be reworded to "this is the only strategy such that there exist no other strategy that has a positive expected score against it."
Here's how to check it.
rock v the strat:
strat choose rock 20%, but score diff is 0 since tie
strat chooses paper 60%, so 0.6 * -1 (for loss)
strat chooses scissors 20%, so 0.2 * 3 (3 for skunk)
add them up ======================================
equals 0
This means that if they choose rock against your strat on average the net score difference is 0.
Repeat for paper: 20% * 1 + 60% * 0 + 20% * -1 = 0
Repeat for scissors: 20% * -3 + 60% * 1 + 20% * 0 = 0
This means that no matter what they choose they cannot beat you on average. You are probably thinking that's nice, but even if they have an inferior strategy, using this will not actually give you an advantage, it will just tie. That is exactly correct. This only tells you how to be unbeatable on average not how to exploit other players tendencies. If you know your opponents tendencies, you can calculate which build you should use against them by just choosing the one that has the highest expected win average. The only time it pays to use a randomized starting build, is when your opponent knows your tendencies and you want to give him no option that on average beats you.
For all this to work we need better estimations of one build versus another. This is where I need your help. You don't have to redo the whole table, if you just say stuff like "I think at the pro level 9 pool has a 40% chance of beating 12 pool." If you are someone awesome like incontrol or midian you don't even need to explain why (other people do though). I do not understand how you can choose a strategy for a game of equal random chance. Are you saying that you understand how to win every rock paper scissors match on average?
On March 22 2008 03:33 flag wrote:
No, im saying I have a strategy for rock paper scissors that no one can beat me more than half the time. That strategy is choose rock 1/3 the time, choose scissors 1/3 of the time, choose paper 1/3 of the time, make this choice randomly.
I'm also saying that this is the only strategy that has this property.
On March 22 2008 05:21 5HITCOMBO wrote:Show nested quote +On March 22 2008 03:33 flag wrote:
No, im saying I have a strategy for rock paper scissors that no one can beat me more than half the time. That strategy is choose rock 1/3 the time, choose scissors 1/3 of the time, choose paper 1/3 of the time, make this choice randomly.
I'm also saying that this is the only strategy that has this property. OH SHIT YOU HAVE A STRATEGY FOR WINNING HALF THE TIME IN ROCK PAPER SCISSORS????
You are essentially making fun of the obviousness of an answer to a question that you asked. That is not a very nice way to treat yourself.
I do not wish for you to harm yourself any more, so please 5hitcombo do not make any future post in this thread.
|
How did you make up those statistics?
|
the one thing I lack in this thing is the probabillity of a person using a certain BO against you.
|
merely as a point of interest, statistics have been gathered from mass rps gaming and it has been undeniably shown that it is NOT a game of chance, as people do NOT pick each choice 33.3333... % of the time.
Beginner players tend to play thus-
Rock is most common, so for their first throw they choose paper, to beat rock.
For non first throw, players tend to subconciously beat their own last throw. So their second would scissors.
I tested this with someone the other and they followed this pattern exactly ;p
As did I to start with, because I forgot lol. Then i remembered this, and won.
Of course more advanced RPS players go deeper into bluffs, double bluffs, etc, etc.
Serious, I'm not even kidding.
|
Calgary25980 Posts
Deleted some posts. I have nothing to add to this thread, but I think it's a worthwhile idea/project. As for how to get meaningful data into it, I have no idea.
|
On March 22 2008 03:33 flag wrote:
No, im saying I have a strategy for rock paper scissors that no one can beat me more than half the time. That strategy is choose rock 1/3 the time, choose scissors 1/3 of the time, choose paper 1/3 of the time, make this choice randomly.
I'm also saying that this is the only strategy that has this property.
Okay, I'm going to post one more thing and leave this thread for good.
Even if you have a strategy like that, you are incorrect in saying that "no one can beat me more than half the time."
No matter what, your opponent can always win if they choose the opposite of what you throw. This is the fault of statistics.
|
On March 22 2008 06:49 ZerG~LegenD wrote:
How did you make up those statistics?
For some of the match ups that occur less frequently I tested them with saarto. For others I watched a bunch of ZvZ replays of JulyZerg and Savior. It really is not very accurate at this point. I plan on attempting a more systematic way to calculate these probabilities soon, but I was also hoping to get some input from others.
On March 22 2008 12:55 MrBobby wrote:
merely as a point of interest, statistics have been gathered from mass rps gaming and it has been undeniably shown that it is NOT a game of chance, as people do NOT pick each choice 33.3333... % of the time.
Beginner players tend to play thus-
Rock is most common, so for their first throw they choose paper, to beat rock.
For non first throw, players tend to subconciously beat their own last throw. So their second would scissors.
I tested this with someone the other and they followed this pattern exactly ;p
As did I to start with, because I forgot lol. Then i remembered this, and won.
Of course more advanced RPS players go deeper into bluffs, double bluffs, etc, etc.
Serious, I'm not even kidding.
I think you might have missed a paragraph from previous post, hopefully this clears it up:
This means that no matter what they choose they cannot beat you on average. You are probably thinking that's nice, but even if they have an inferior strategy, using this will not actually give you an advantage, it will just tie. That is exactly correct. This only tells you how to be unbeatable on average not how to exploit other players tendencies. If you know your opponents tendencies, you can calculate which build you should use against them by just choosing the one that has the highest expected win average. The only time it pays to use a randomized starting build, is when your opponent knows your tendencies and you want to give him no option that on average beats you.
On March 22 2008 14:33 Chill wrote:
Deleted some posts. I have nothing to add to this thread, but I think it's a worthwhile idea/project. As for how to get meaningful data into it, I have no idea.
Thanks
|
I am very confused. I'm gonna say how I interpret this, and you correct me.
OK, so my understanding is, you start with a table with every mu vs. every other mu. You can make up (from watching games, pull numbers our of your ass, any method you wish) a table of probabilities (ie. you randomly say that 4 pool has 10% chance of beating 5 pool). You then make up another table of how likely the opponent is going to use a certain strat (ie. 10% of going 4 pool, 60% chance of going 9 pool, etc.). You then crank out some math based on positions and these two tables to determine which build "overall" will give you the most amount of wins over the long run.
Is my understanding correct?
Although this table won't be that useful, as you would have to recompute all of the tables for each player and each map, and even then the probabilities are very rough ballparks that won't decide how you play the game anyway. Though I do think it's pretty interesting, and things don't need to be useful to be interesting.
Just wondering, is your program as flexible as I think it is? Like, you can change any of the percentages around, you can add bo's (thus we can apply this to any mu we want)? How hard would it be (aside from redoing all the numbers and charts) to say, change the map from Python to Hunters (changing from 4 players to 8 players would screw up some of the scouting probabilities, no?)
|
Your understanding is basically correct except there is not a separate table for how likely your opponent is to use each strat, instead it uses the results for that.
The program doesn't have anything about starcraft in it, it just takes input of a matrix which has the percent win for each match up. And outputs what percent of the time to choose each. So for each match up, map, etc all the percentages would change and right now that would have to be done manually.
I have an idea to generate the percentages automatically using this method:
Simulate the builds, for every x seconds a player has lings in the other's base while they have none, give a drone kill. Once both players have lings it is now just a calculation of economic strength versus each other, this can be measured in how many seconds it would take one player to catch up economically to the other. Now the percent chance of win is solely a function of economic difference. This is not perfect, real starcraft is more complicated than this in many ways. Some of these ways could be accounted for but there are really too many. Two that I have in mind now are a small bonus of some sort for the aggressor. And some sort of bonus for getting gas up earlier.
This method will not be perfect but it will be consistent for all build orders at least. It could also be changed to other maps easily because only variables are time to get to them and a table of how many minerals per second per drone you get.
|
5hitcombo, I always thought of you as a pretty intelligent poster, and I am unfortunately annoyed by your blatant stupidity in this thread.
flag, I really appreciate your work into this, even if there are no major revelations that come from this. I think the biggest reason is because each individual player has their own tendencies, and the most important thing is to adapt to that particular opponent. Obviously a 1/3 1/3 1/3 strategy in RPS is optimal according to game theory, but since people do different strategies, there are counter strategies. Therefore, you make a small mistake of deviating from "optimal play" in order to take advantage of your opponent's "larger mistakes". (this ties into fundamental theory of poker, etc..)
Simple Starcraft example would be if someone constantly 4 pools, then you would 5-9 pool every game, despite this not being the "optimal strategy" in your table. If I play against Julyzerg, it is therefore more intimidating than a more 'conventional' zerg, seeing how he is capable of very rarely 4pooling, making his potential "build range" wide.
Bottom line is, while it is good to try to seek perfection, true perfection is knowing your opponents tendencies and making your 'build order range' as wide as possible (you are capable of anything), while subtly adapting to exploit your opponent.
|
Well done. I've been thinking of doing this as well for a long time, but i couldnt really convince myself of the usefulness. :/ anyways, im happy to se it done. 
For those that want to go more into detail, the solution is refered to as a "Nash equilibrium" and you can read more about it on wiki. Correct me if this is not how you've done flag, but Im pretty sure this is how you did it.
http://en.wikipedia.org/wiki/Nash_equilibrium
Reading just the first paragraph a few times should do the trick.
Note, that while, as stated already, this isnt useful in the case of a public game on bnet, it is highly useful for progamers. Difference is that your oponent in progaming will have access to all opening you've ever done. The (mathematically simpler) case of random pubby was cover by chill a while ago iirc.
So you basically just made up the winning percentages yourself? Or could you explain a bit closer exactly how you got them. I think the percentages really need to be looked over for this to be anywhere close to reality.
Im not a good player myself, but isnt 42% for 12 pool > 12 hatch a bit high? If you'd down that percentage a bit, 12 hatch would be more viable, so itd be done more, which would make 9 pool more viable, giving results closer to reality.
Im not sure. I would love it if this would produce useful infomration, but I think you have to find a better way of getting the percentages for this to be useful.  Like statistics from every ZvZ ever played by all progamers... So if someone got a year or two to spare for just watching ZvZs and taking statistics, we could get this going. 
|
|
|
|
|
|
EPT
