EPTOn October 15 2016 05:46 jalstar wrote:
There was a TLPD page that had BW map balance using an algorithm that took player skill into consideration, similar to what people are suggesting here. You had to have the URL though, there was no link to it. I think it's gone now.
If P4ndemik or PoP still post they might know what I'm talking about.
edit: It was at http://www.teamliquid.net/tlpd/korean/maps/balance_table.php and it's gone now.
There was a TLPD page that had BW map balance using an algorithm that took player skill into consideration, similar to what people are suggesting here. You had to have the URL though, there was no link to it. I think it's gone now.
If P4ndemik or PoP still post they might know what I'm talking about.
edit: It was at http://www.teamliquid.net/tlpd/korean/maps/balance_table.php and it's gone now.
Doing it through TLPD would be ideal, because that was you'd have all the data at your disposal immediately, particularly things it could remedy:
1) Chess ELO works that for a 400 point difference, you have a 10/11 chance of winning. I don't know the exact Liquipedia formula, but you could take the instantaneous ELO at every point and combine it into win statistics for every game. This would give a weighted value of win rate, the ELO used could be some combination of vs match-up ELO and total ELO. This would remove the assumption that players of all races are equal in skill level, and it'd take into account good players being bad players on an imbalanced map (I think with sample sizes this large, this effect would be almost negligible, but many of your disagree).
2) It could take into account the number of games played by each race in different timeframes. Currently my assumption was that equal number of games by all races on all maps (I think when looking at large enough time frames, it's reasonably accurate)... But the issue this causes is that if there's 80 TvT's played and 40 PvP's, my algorithm reads this as hey, it's Terran favoured, as Protoss are trying to dodge playing on this map (particularly in PL). However it's possible that there are only half as many protoss as terran, and hence that's not the case. It'd fix that bias, I think this could/would definitely provide a benefit to the current system, particularly at the upper end of balance
3) Instead of giving the map one single multiplier for a year of maximum popularity, it's balance could be calculated at a few different periods (or give a slightly larger weighting to the later on areas to calculate the winrate). For example, games in 2012 would count as 1.5 games, so if it goes 30-30 in 2012, but 30-20 in 2008 (which would have an arbitrary value of 1), the weighted winrate would be 53.57% vs 54.55%.
4) Potentially give a different weighting to different tournaments depending on what VETO system they have, since no-vetoes make it that the amount of mirrors played isn't as significant as in say PL, where teams will never put out a terran on a map where Terran does poorly.
Some things could never be taken into account, such as the manner in which the game was won, among other things like how good a player really is, for example iloveoov's dominance in numbers never looked all that impressive to me, same thing about say Conor McGregor in UFC, that's more about style and aura that's given off.
Anyway, a statistical analysis is always imperfect, the world has too many variables and too much chaos, but often it's the best we can do, so making the best model possible and analysing it might give the most meaningful results for the discussion.
edit: Not enough data in TLPD to gather data for winrate vs time, but it'd be neat to see. I suppose you could add it into the balance model, but not necessarily sure what's more balanced - 100% winrate in early game and 0% winrate in lategame, versus 50-50 in both (assuming same number of games ending in early and lategame). I quite like the dynamic that you just need to weather the storm and then it's smooth sailing (zerg trying to get out defilers out in time for example), particularly common in Dota. Don't want to make too many subgroups of the data though, as outside of the 20~ big maps, the sample sizes for the rest aren't large.
