The Curious Case of Treant Protector

October 29, 2012

Too distracted by Centaur to get anything important done. It’s a shame the API is down because this last week would have some pretty fun stats. Can only get bits and pieces from Dotabuff. Did you know that Centaur singlehandedly improved Blademail’s win rate by 10%. It’s true.

Anyway, Treant Protector. Treant in 6.74 was one of the trio of near 60% win rates, along with Lycan and Ursa. Lycan and Ursa are pretty self explanatory. Fairly mindless right clickers, jungle capable, massive built-in steroids and relatively gear independent, capable of putting Roshan into play very, very early, etc. Treant was weird though because as successful as he was at pubbing no one seemed to notice. He went completely ignored in competitive games, and he wasn’t terribly popular outside of them either. No one really knew why he was successful, and no one really seemed to care either.

But I found it curious. Treant being successful in pub isn’t too surprising. His closest analogue is Tidehunter, who typically maintained a ~55% public win rate. The complication with this is that Tidehunter’s win rate is generally attributed to his Ravage. Treant’s ult was at the time widely considered inferior in all possible comparisons.

The one clue I came across is that Treant’s win rate definitely deteriorated in the upper skill brackets. Whatever it was that he had going for him was something that worked disproportionately well against weaker players. Treant wasn’t alone in this. Omniknight and Death Prophet exhibited similar, but weaker trends, probably due to the fact that players in the lower end of the skill distribution don’t know how to deal their ults. But Treant’s ult was similar to Tidehunter’s, and Tidehunter didn’t experience the same kind of fall off. Clearly something else in Treant’s kit was bumping him up to top 3 status.

The two clear possibilities was Treant’s invisibility and his passive. Invisibility seemed unlikely. If it was boosting his win rate it would be doing it in an environment with teamwork and communication, which is pretty much the opposite of the games he was actually winning. The passive was the much more likely possibility, and honestly the much more intriguing one. It’s hard to believe that some small global healing and armor half of the time could be one of the most game-changing abilities in low level play. But one aspect to consider is that the passive did apply to turrets, so perhaps it was an incredibly successful anti-push ability in an environment that rarely featured coordinated pushing.

Then 6.75 came along. Treant saw two major changes. His ult lost its damage, but received a duration boost and reduced cooldown in exchange, and his passive was revamped to an active, single target global heal/refraction. As a result Treant’s near 60% win rate dropped all the way to the lower mid 40s.

I think the net effect of Treant’s ult changes was a slight nerf in public play, but the passive revamp was the much larger change and reinforces the hypothesis that his old passive was surprisingly one of the most broken abilities in the game in low level play. In light of this, I wonder how successful a player in the normal bracket could be if they specialized in Treant, building a quick mana boots, and then just literally using his global on (its now 15 second) cooldown. Mana boots alone should be able to cancel the mana cost (15 sec cooldown, 4 uses per minute at 25 mana per use, 100 mana a minute, Arcanes restore 110 mana every 55 seconds). You should be able to get more out of it than the old passive, and it would probably be decent practice for map awareness. On a 15 second cooldown you can basically just guess if a fight is about to break out and barely be losing anything if you’re wrong. If a turret is damaged, you can just chain cast on it until it’s back to full. And always cast it on a turret taking damage either before or after glyphing. Theoretically this could even be more broken than the old Treant, and more than easy enough to pull off as a solo suicide in lower level games.

That does it for today. Hopefully I’ll spend some time in the next day or two figuring out the best way to integrate the chartiest charts you ever charted into this thing. Until then.

Advertisements

Hero Usage and Win Rates in the Three Skill Brackets

October 24, 2012

One of the first big things I hoped to accomplish with the API was use it to examine how hero usage changes at different skill levels. Dotabuff.com has done an extremely good job establishing usage and win rates in the aggregate, but unlike the API, their match collection methodology is incapable of filtering by skill level.

While the API was still working, I grabbed three samples of 10,000 games from each of the three skill brackets. They’re smaller than I would like them to be, but they’re mostly workable samples. The High and Very High samples are collected from various times over a period of about a week. The Normal sample, unfortunately, spans a period of about two months due to a API bug that I had to work around towards the end of its lifespan. Because of this and the fact that the Normal bracket is by far the largest (making my 10,000 Normal sample the least representative) I prefer to use the 3-day stats from Dotabuff’s predecessor as a Normal estimate whenever possible.

The results of this project can be found here: https://docs.google.com/spreadsheet/ccc?key=0AoNi7mtSTYNzdFdyUUJ3X1ppQ3VxVFhVVXc3WEp2Z1E

It’s still fairly basic, and I hope to have an improved version ready for when the API gets fixed and I can expand my samples. This is purely 6.74 data.
To do your own sorting you must save a local copy. It’s a little inconvenient unfortunately, but I don’t know of a better way to handle it at the moment.
It’s also important to keep in mind that the sample size is not as large as it ought to be. For usage rates, the error level is relatively acceptable, but the small sample size can wreak havoc on the accuracy of the win rates. Take any extreme result in win rates with a grain of salt.

Now that that’s out of the way, onto some of the interesting trends.

Comparing the Dotabuff usage rates to the Very High usage rates we have these heroes seeing the biggest increase:

15.59% Rubick
14.84% Windrunner
14.18% Nature’s Prophet
13.29% Invoker
7.64% Leshrac
7.59% Queen of Pain
7.32% Shadow Demon
7.18% Mirana
7.04% Enigma
5.12% Ancient Apparition
4.86% Sand King
4.74% Storm Spirit
4.60% Crystal Maiden
4.58% Morphling
4.48% Dark Seer
4.21% Disruptor
3.55% Wisp
3.46% Chen
3.38% Tinker
3.11% Venomancer
2.74% Puck
2.25% Lone Druid
2.12% Earthshaker

There are two big trends I see from this.

1. Very High games are much more likely to pick dedicated suicide laners. Windrunner, Queen of Pain, Mirana, Dark Seer, and Puck are all quite capable in the suicide lane, and even Nature’s Prophet, Enigma, and Lone Druid can have some success there.
2. There is much more of an emphasis on farm independent heroes. The only truly gold dependent heroes in the list are Morphling, Storm Spirit, and Lone Druid, all of whom have significant skill checks that might scare away players in the Normal bracket.

On the other end of things:

-11.86% Riki
-10.65% Drow Ranger
-8.94% Huskar
-8.18% Bloodseeker
-8.10% Phantom Assassin
-7.93% Sniper
-6.76% Ursa
-6.01% Skeleton King
-5.80% Viper
-5.54% Spiritbreaker
-4.66% Bounty Hunter
-4.19% Zeus
-3.94% Dragon Knight
-3.84% Axe
-3.78% Night Stalker
-3.78% Nyx Assassin
-3.70% Keeper of the Light
-3.38% Faceless Void
-3.07% Omniknight
-2.96% Sven
-2.82% Juggernaut
-2.82% Phantom Lancer
-2.82% Death Prophet
-2.67% Doom Bringer
-2.66% Slardar
-2.62% Luna
-2.44% Clinkz
-2.23% Razor
-2.20% Spectre

What’s noteworthy here is how so much of the usage decline is monopolized by carries. Only Zeus, Axe, Nightstalker, Nyx, Keeper, Omniknight, and Death Prophet stand out as being relatively gold independent, or closer to what is commonly referred to as semi-carries.

Moving on to win rates increases:

7.46% Chen
6.51% Pugna
5.89% Templar Assassin
5.31% Batrider
5.13% Phantom Lancer
5.01% Clinkz
4.34% Wisp
4.16% Bounty Hunter
4.04% Visage
3.68% Enigma
3.41% Storm Spirit
2.99% Faceless Void
2.90% Morphling
2.28% Beastmaster
2.04% Invoker
2.01% Ancient Apparition

Win rate increases are tied most strongly to characters with a high skill cap or that require successful early aggression. Chen and Visage are obviously quite micro intensive. Invoker needs no explanation. Enigma, Faceless Void, and Ancient Apparition are all dependent on hitting big ults to pull their weight. Wisp requires a hell of a lot of team coordination to play to the best of its abilities. Templar Assassin, Batrider, Clinkz, Bounty Hunter, Storm Spirit, and Beastmaster are all dependent on using their kits for successful early and mid-game aggression. Morphling and Pugna are both relatively complicated characters to play successfully. Phantom Lancer is the one oddity on this list.

On the losing end:

-7.60% Treant Protector
-6.32% Outworld Destroyer
-4.86% Drow Ranger
-4.86% Skeleton King
-4.38% Axe
-4.36% Undying
-4.18% Sniper
-4.02% Jakiro
-4.01% Dark Seer
-3.83% Death Prophet
-3.45% Crystal Maiden
-3.39% Keeper of the Light
-3.27% Omniknight
-3.08% Earthshaker
-3.04% Disruptor
-3.03% Gyrocopter
-3.02% Silencer

This is a little less clearcut, and the perils of the small sample size might play into here. Treant Protector being the biggest loser is something I plan to expand upon in a future post. Aside from that, I’ll simply leave this category to your interpretation.


Update to Dota’s Skill Bracket Size

October 18, 2012

I ran with 80/16.5/3.5 as a size estimate because I couldn’t find a statistical structure that better explained what was going on.  Turns out I was a bit blind.

My new theory is that the skill brackets are determined using the standard deviation of the rating distribution.  The slice between the mean and one standard deviation above the mean is 34.1% of the distribution.  The slice below the mean is 50%.  34.1+50 is 84.1, and the difference between 84 and my estimate of 80.5 can be explained by the High and Very High brackets having players that are, on average, more active than the Normal bracket.

This theory puts the High bracket cutoff at 2 standard deviations above the mean, or 13.6% of the population.  The remaining 2.3% is the Very High bracket.