Designated Balance Discussion Thread - Page 1031

On July 12 2014 06:10 Big J wrote:
@Grumbels: Sounds pretty much like what you'd expect. Only a few things I find fishy:
-) the increased Terran vs Protoss winrate in your example when favoring T over Z. I'd expect the whole thing the other way around.
-) not sure how you'd reach the 50% balance in the longrun, given that this is a matter of qualification format as well. Also I think you wont reach 50, but rather something in between the imbalance and 50, depending on the qualification format.

Anyways, great initiative. I will try to replicate your experiment tomorrow.

On July 09 2014 03:07 Grumbels wrote:

Unit            Supply  Minerals Gas    Build Time
Battlecruiser	 6	 400	 300	 90
Carrier	         6	 350	 250	 120
Tempest	         4	 300	 200	 60
Colossus	 6	 300	 200	 75

Note that protoss units build quicker in practice because of chronoboost.

There are a number of reasons for the colossus and tempest to see more use:

• The battlecruiser and carrier are much more expensive and slower to manufacture than even the colossus and especially the tempest. This leads to huge opportunity and infrastructure costs, which makes getting only one of those units hardly worthwhile.
  
• Colossus attack upgrades are also more accessible.
  
• The colossus and tempest also fulfill necessary roles (ground control vs light units / deterrence vs zerg late-game), so you are more often immediately rewarded for building them, despite the set-up costs.
  

I think looking at these factors the obvious conclusion should be that battlecruisers and carriers have a quite different function than other T3 units (that of late-game ultimate army), one which is probably not necessary in this game, and not feasible at a pro level where timings are so strict. If you want them to see use you have to make them weaker and cut down on build time and cost so that they're more in line with the tempest. It would probably also help to give them a well-defined role, for instance like carriers had in Brood War.

There is also the question of critical mass: brood lords benefit from broodlings obstructing, and therefore become exponentially more powerful in higher numbers as you can no longer shoot down all the broodlings in time to attack their sires. I think that's why you don't really see brood lords any more: it takes too long to get to critical mass without the addition of the old fungal, and you need enough brood lords that they become vulnerable to anti-air. Nevertheless, they do fulfill a certain function and can effectively space out ground armies once they reach critical mass. I don't know if exponential increase in strength for such units is specifically harmful, an argument can be made that it simply denotes the point when the unit can actually do its job -- brood lords can finally succeed at space control, and will still have numerous strategic weaknesses to offset this strength.

Battlecruisers and carriers don't really reach critical mass, but that's also because they can't really do anything so it becomes pointless to talk about critical mass. You could say that the units would be more worthwhile to build if even a few of them could change the battle, but without a well defined role they will still feel a bit pointless imo. Though if the units are too weak they will never be strong at anything obviously, and have no function in the army, so it goes both ways. (sorry if this logic is a bit vague, slightlarge headache.. )

So yeah, reduce costs...

On July 12 2014 06:26 Grumbels wrote:

Sorry, that's a typo. I meant the other way around.

I thought it would be fun to use "if you make protoss overpowered" as the main example everywhere specifically to annoy people, but then I thought it would undermine the point I was trying to make, so I changed it to terran. But then I forgot to change it elsewhere.

On July 12 2014 03:54 Grumbels wrote:
I made a tournament simulation for fun to check some things:
Given a randomly distributed bracket with p, t, z of equal range of skill,
- If you favor, for instance, terran in its match-up vs zerg, then zerg vs protoss will become zerg favored and protoss vs terran will become protoss favored. Although these effects are quite small compared to the advantage terran has vs zerg. Another interesting detail is that if you compare the results of favoring terran vs either only zerg, or both zerg and protoss, that you will find that TvZ is more broken in the former case just by looking at the win rate.
- If you make mirror match-ups more 'coinflippy', for instance protoss vs protoss, then protoss win rates will fall down in other match-ups as more weaker protoss players are advancing. Similarly, if you make, say, terran vs terran more skill-intensive, terran will become favored vs zerg and protoss.
- People say that only the balance at the top matters, but if you add additional good, if not great players of one race, they will produce more tournament winners, sort of depending on the values you choose. I think it matters if one race is "easier to play" than another race for tournament results even if top players are still equally matched. Obviously this is because the better player doesn't always win and as a race you're going to do well if you have more players near the top.
- On the other hand, adding just one very strong player for a race can singlehandedly save their win rate, partly because a good player will play more ranked games than a weak player and have more influence on the win rates.
- By manipulating the bracket you can give one race an almost 50% chance of winning the tournament instead of 33%. Not to encourage tournament organizers to do this, of course.

Actually, I tried looking for the effect where the win rates will self-correct to 50% over time, but I couldn't find it. Anyone knows why? (I'm just running single elimination brackets and checking win rates per match-up per round after giving one race a global rating boost. I suspected the win rates would be closer to 50 every subsequent round, but in fact the opposite happened) I think I'm misunderstanding the concept somehow, or maybe there's a problem with the simulation.

Anyway, I think that these results reinforce the idea that balance is more complex than looking purely at win rates. It's just some quick tests though, don't know how representative the results are or if I did something really dumb.

Any thoughts?

On July 12 2014 13:24 ChristianS wrote:


Edit: I should issue a correction for that blog: I said that an individual player plays three different matchups in SC2. Scarlett apparently plays four.

On July 12 2014 23:01 Big J wrote:


So, without being quite finished with my GSL simulator, my first testings seem to produce the asymptotic 50% winrates.
But I'm not sure how much work I can put into that project until middle of the next week, that's why I'm giving a quick update on it long before I can post details or graphs, TT

Edit: Aaaaaaaaaaaand, mistake identified and its gone.

On July 13 2014 02:58 Grumbels wrote:

I think it makes sense though.

With my simulation I was singling out the top players (i.e. the ones advancing in the tournament) and if you strengthen one race it will not only have better representation but stronger top players. However, I suppose that if you had the middling players face each other in endless consolation matches, that you would start to replicate the predicted effects of asymptotic win rates that you might also see in the ladder.

I don't know if it's true, but I imagine that code A qualifiers will be closer to 50% than Code S/A in times of imbalance. And that GM/Masters winrates might be more revealing of imbalance than Diamond and below.

Which in my opinion possibly makes it dubious to state without proof that win rates from, say, the Korean scene are meaningless because of this tendency to go to 50%. I'm sure it could be true, but it probably depends on your sample, so in that case it's not categorically true, and then we'd really need someone with more statistical expertise to inform us.

Actually, I tried looking for the effect where the win rates will self-correct to 50% over time, but I couldn't find it. Anyone knows why? (I'm just running single elimination brackets and checking win rates per match-up per round after giving one race a global rating boost. I suspected the win rates would be closer to 50 every subsequent round, but in fact the opposite happened) I think I'm misunderstanding the concept somehow, or maybe there's a problem with the simulation.

- On the other hand, adding just one very strong player for a race can singlehandedly save their win rate, partly because a good player will play more ranked games than a weak player and have more influence on the win rates.

If you make mirror match-ups more 'coinflippy', for instance protoss vs protoss, then protoss win rates will fall down in other match-ups as more weaker protoss players are advancing. Similarly, if you make, say, terran vs terran more skill-intensive, terran will become favored vs zerg and protoss.

Initial1: Create a playerbase (e.g. 128 or 1024players), that have a skillindicator and a raceindicator
Initial2: choose 32players randomly from the playerbase and let them play a Code S tournament
Step1: choose 48new players from the playerbase and let those +the bottom 16play a Code A tournament. The next Code S are the top16 of this tournament and the top16 of the last Code S tournament.

- create a player pool of 256 players
- - using Elo ratings to denote skill
- - evenly distributed skill & race
- for n times
- - put them into a 256 man single elimination tournament with random seeding
- - run the tournament and record the win rates per round and count winners
- get results.

Note that, unlike what Grumbels was talking about and what I thought I'd see on a smaller scale, there seems to be hardly any effect on the other matchups.

I have been working with a few calibrations for skilllevel. Currently, I'm creating Gaussian distributed values around 200 with a standard deviation of 50. So a standard player has roughly a value between 150 and 250, with extremes to 350 or 50 (and a mainly theoretical cap at 400 and 0).
I realize this isn't the best calibration though, because given a player with 250 skill playing against one at 150skill, this merely produces a 62.5% winrate. Basically we could be talking about somewhat coinflippy matchups.
I will try to work with this a little more, but given that results with other calibrations haven't been very different, my guess is that it does not matter that much for these grand scheme pictures.

On July 13 2014 21:08 Hider wrote:


I don't understand this proces.

Why do you choose 32 players randomly from the database and let them play a code s tournament?

Here is what I would do:

Have a database of like 1xxx players with a random distribution of skill level points.
Probably of winning should be based on skill level + race strenght.

The top 128 (this is based on skill level + race strenght) gets into a Code A win or get knocked out tournament
The top 32 of in the Code A tournament qualifies for a Code S tournament. The losers don't play any more games.
Code S tournament is also win or get knocked out.

In this proces you make sure that there will be fewer players of the UP race who gets recorded games, which means the win/rates of the UP race gets much closer to 50/50. With your methodology the UP race gets a solid amount of games regardless of their succes.

On July 13 2014 21:31 submarine wrote:


I have the feeling that player skill is not valued enough in your win formula.
A 350 player has a winchange of 50% against someone with just 210 if he plays the weaker race.

Maybe try it with a lower imbalance setting?

EDIT:
Do you actually see a accumulation of high skilled players in your model?

On July 13 2014 21:39 Grumbels wrote:

The red line seems a bit below 50% though, it seems significant enough. In my testing the effect would be about ten times smaller than the initial imbalance, so it's quite small but still noticeable. In comparison, the effect of making certain match-ups more / less 'skillful' were stronger.

(of course we're using different tournament formats and yours seems more realistic, which is why I'm hesitant to state the exact values)

On July 13 2014 21:39 Grumbels wrote:

Should the calibration matter at all? Also, does it matter that you're using Bo5 instead of Bo1? I don't think it should have any effect on the results. No matter how small the rating difference, it will never be at exactly 50% and therefore will always show up given enough simulations, and no matter how high it will never be at exactly 100% either .

In the consecutive tournaments, by means of the qualification progress, I'm making sure that the weaker players and the players of the UP race drop out.

On July 13 2014 22:33 Grumbels wrote:
@Hider, those values are completely arbitrary though.