Where do you think the APM figures in the AlphaStar paper came from? They are the real APM values from about 900,000 ladder replays. They are the “in game” APM. It’s like I am talking with a vaccine denier for crying out loud. What’s next? Perhaps you’d like to propose a conspiracy by the Deepmind team to fake their APM charts? Golly, you got me. I have no answer to that one /s.
LoL there’s more flat earth believers than there is SC players. I’m not one of them, but hey I really don’t care too much about SC 2 any more. There’s too many crap talkers online so I kind of lost interest. Fun game, trashy talk!! But there’s still lots of cool players online so it’s not a total loss, the trash talkers ruin it for everyone… so high APM low APM it doesn’t really matter to me, but I’m glad I could get a nice thread going 
APM = EPM + redundant
So why would this not be the case? A lot more correct actions are by gm you cant say bronze makes equal number of correct actions.
And they are right. How do I know? Because I have been and trained with in clans of Masters plenty to know how they play. The spammers start from D2 or D1 up but it is covered by some skill… I mean isnt as bad as D3 and lower
I asked you to explain that picture https://i.imgur.com/lKowzyx.png, and to provide the links to your data, not your paint pictures. What do I get in answer? Another number out of your head? There are 900000 replays. Where do you take that number from? There are 0 links in google which mention that number, same goes for 900k.
But hey, everyone who does not believe a random guy on the forum, who makes paint pictures and present them as a proof is vaccine denier…
I told you it’s from the alphastar paper. They even include instructions on how to mass-mine replays from the Battle.net API.
Supervised learning
Each agent is initially trained through supervised learning on replays to imitate human actions. Supervised learning is used both to initialize the agent and to maintain diverse exploration56. Because of this, the primary goal is to produce a diverse policy that captures StarCraft’s complexities.
We use a dataset of 971,000 replays played on StarCraft II versions 4.8.2 to 4.8.6 by players with MMR scores (Blizzard’s metric, similar to Elo) greater than 3,500, that is, from the top 22% of players. Instructions for downloading replays can be found at [link]. The observations and actions are returned by the game’s raw interface (Extended Data Tables1, 2). We train one policy for each race, with the same architecture as the one used during reinforcement learning.
From each replay, we extract a statistic z that encodes each player’s build order, defined as the first 20 constructed buildings and units, and cumulative statistics, defined as the units, buildings, effects, and upgrades that were present during a game. We condition the policy on z in both supervised and reinforcement learning, and in supervised learning we set z to zero 10% of the time.
To train the policy, at each step we input the current observations and output a probability distribution over each action argument (Extended Data Table2). For these arguments, we compute the KL divergence between human actions and the policy’s outputs, and apply updates using the Adam optimizer57. We also apply L2 regularization58. The pseudocode of the supervised training algorithm can be found inSupplementary Data, Pseudocode.
We further fine-tune the policy using only winning replays with MMR above 6,200 (16,000 games). Fine-tuning improved the win rate against the built-in elite bot from 87% to 96% in Protoss versus Protossgames. The fine-tuned supervised agents were rated at 3,947MMR for Terran, 3,607MMR for Protoss and 3,544 MMR for Zerg. They are capable of building all units in the game, and are qualitatively diverse from game to game (Extended Data Fig.4).
They also had to do a significant amount of research on APM, since they needed to impose realistic / human-like APM constraints. The APM probability density functions for each race are defined on page 3. You can’t miss it. There are no “camelback” distributions to be found. APM spam doesn’t exist to any meaningful degree.
My man, you couldn’t tell that the axes on the charts were mislabeled. That’s what I am dealing with here. You don’t even know what you are looking at. Tip of the day: always include a small error which allows you to gauge if the people you are talking to know what they are talking about. I do this all the time: “Hey, did you know that Trump will have to testify before the house committee?” and if they respond “Oh yeah he will get rekt” then you know they’re a dunce. The house committee will dissolve faster than any legal challenge can resolve, ergo he doesn’t have to testify as long as he has a lawyer file a challenge to it in court. But, if they respond with “That’s literally not how it works” then you can feel happy inside and say “Darn, I almost had you!”.
There was a forum troll called “nomufftotuff” who caused me to develop this method. The guy was an absolute nut. He demanded to see my algorithms for analyzing Aligulac’s database and gave all these various reasons why I was wrong. Except, I had provided the math and it had an error which he didn’t notice while he rambled on and on about all this totally irrelevant nonsense. So, that’s your pro tip for how to deal with trolls in general. If someone knows their stuff, they will call BS on you instantly the moment you say something incorrect. When you include the error and nobody says anything, you know any conversation with these people is a waste of time.
Now if you don’t mind me, I am going to go stare at my biceps in the mirror. 
By the way, I am not the only person to have published these findings. Some Blizzard employees did a slideshow presentation wherein they said the correlation between APM and win-rate was 0.63 (if memory serves). There are other papers (link below) which show the correlation between MMR and APM to be 0.65. I once calculated how much APM can vary on a per-individual basis for the corr(Winrate, APM) == 0.65 to be true. In 95% of cases, a player’s APM will vary from what’s predicted by their win-rate by no more than 26 APM. It’s simply a fact that APM is the majority of the skill in the game.
https://iccm-conference.neocities.org/2020/papers/Contribution_287_final.pdf
Side note: at the time I made that calculation, I didn’t have access to mass replays for data mining. I calculated that off of the figures from the slideshow, an average APM table for each league, and the assumption that APM was normally distributed. Well, this paper confirms my calculations (congats to past batz) and shows my assumptions were justified (congrats again to past batz). Source:
I’ll add a quick sub-note to this. This is calculated for players in masters league under the assumption that corr(MMR, APM+scattering)=0.65. When calculated for the ladder as a whole, the result is Scattering ~ N(0, 65). To contrast my simulation against the data in the paper I referred to previously:
https://i.imgur.com/QfIlmxx.png
As you can see, it’s spot on. 
I’ll add another sub-note. Because MMR is a very dynamic system and players’ mmr values are always fluctuating up and down, it’s likely the vast majority of the variability in this correlation is due to the instability of MMR and not APM. I don’t have any figures on the variability of MMR so I don’t know exactly how much to attribute to it but we do know the correlation between average MMR and average APM is certainly > 0.65. It’s probably something like 0.8.
Now care to explain to me please that picture from the link you provided https://imgur.com/a/B8MB2Cm because what I see here is that some guys with 3k mmr have 300 apm, just like I said, while some guys in masters have less than 100.
I already imagine myself in the college telling the professor, that I don’t have any idea of variability or correlation, so I would use a random number for it. And I would go even further: I would neglect the provided graphics, which do show, that players with higher APM can have less MMR, I would fill it with random numbers to prove, that the graphics provided by researches are wrong and do not display their real APM to fit my agenda.
Why don’t we quote the researchers?
Webber et al. (2010) studied players of various skill levels
and found those with consistently higher APM usually per-
form better in RTS games such as StarCraft; their analysis
concludes this is due to experts encoding ballistic action se-
quences.
Translation, high APM players are better players.
Huang et al.(2017) found that some novice and most ex-
perts produce excessive APM during the first two minutes of
a game; those interviewed associated it with a warm up (not
unlike sports games); however, a distinguishing feature be-
tween novice and expert is the consistency of APM through
the rest of the match – experts rarely decline in APM, whereas
novices drop off during periods of intense or confusing states.
Further, they also found that experts are more likely to use
unit groups for production buildings (vs mobile units), rebind
unit groups on-the-fly, and retain consistent habits regardless
of game state
Translation, APM spam only happens to fill spare time in the boring early game moments, and newbs APM drops off as the game becomes more complex and confusing, ergo they have to spend more time thinking about their actions before making them.
In Figure 2 we see a moderately positive correlation ( r =
0.65) between player APM and MMR extracted from each
SC2 replay file metadata. This indicates a relationship be-
tween the two, with higher average APM likely being the re-
sult of player skill level, rather than the cause of it.
…
The increase in mean APM is somewhat expected, as play-
ers familiar with the game and action sequences will naturally
initiate, queue up, and respond to in-game actions on the fly
Translation, experts have the correct answers memorized so they don’t have to spend time thinking about the right answers.
Now, let’s contrast that against my first post in the thread:
As you can see, I am in the great company of these six fine researchers. APM is roughly equivalent to skill especially if combined with other metrics like SPM and number of hotkeys used, and APM spam is rare / only happens in the early game. But, some forum troll said I was wrong because he was confused on how to read a chart. Lmao. However will I combat such a formidable argument? I feel all alone. Why doesn’t ANYBODY share in my opinions? 
Yes, that’s unironically exactly what you would tell your professor. I would simply quote the various researchers like I just did.
In case it isn’t clear what I am doing here, let’s let Papa Roach do the explaining:
What? You finally acknowledge that APM spam does exist?
Keep up, and you might finally understand that picture https://imgur.com/a/B8MB2Cm, where the line shows, that with each league the APM raises up, but in MANY CASES higher in-game apm =/= higher skill.
Sad that even researchers quotes are too hard for you to understand. Otherwise you would be able to understand what:
studied players of various skill levels and found those with consistently higher apm USUALLY per-
form better in RTS games such as StarCraft
actually mean. The word usually is there for the reason. For a reason I said like 5 times already.
My opinion was almost verbatim of that of the researchers. You’ve got nothing here, my man. You can ram your face against this brick wall but it won’t budge. Players have high APM because they understand the game well and don’t have to spend as much time thinking about what actions are the correct actions. They have the correct actions memorized. APM is equivalent to skill.
Or because they click to much, just like these diamond guys with 300 apm https://imgur.com/a/B8MB2Cm
To quote the researchers:
Huang et al.(2017) found that some novice and most ex-
perts produce excessive APM during the first two minutes of
a game; those interviewed associated it with a warm up (not
unlike sports games); however, a distinguishing feature be-
tween novice and expert is the consistency of APM through
the rest of the match – experts rarely decline in APM, whereas
novices drop off during periods of intense or confusing states.
Further, they also found that experts are more likely to use
unit groups for production buildings (vs mobile units), rebind
unit groups on-the-fly, and retain consistent habits regardless
of game state
APM spam doesn’t exist outside of the early game when nothing is happening. APM is equivalent to skill especially when combined with other metrics like the number of hotkeys used.
Your own quote, which literally says, that it is USUALLY the case, means it is not always the case and provided graphs literally display that. Each of your graph, stat or quote only makes it worse for you.
That’s why they delve deeper into their APM analysis. There is more to APM than just the average APM. You can also look at how many hotkeys are used to create that APM, how many screen movements, etc. These are all bundled up in the APM metric. Doing a deeper dive shows even broader APM differences between players since higher skilled players use more hotkeys and move their screen more. APM=skill.
It’s really weird that player from GM doesn’t know, that moving the camera, using camera hotkeys does not count in in-game APM. What counts is giving actions to the units, choosing the units with mouse or a hotkey.
APM is the least granular input to the game. If you hit a button on your keyboard or click your mouse, it’s APM.
Tell it to the game, because the game counts only commands given to the units, or selecting the unit with hotkey or mouse. You can move your mouse or camera as much as you want, no one counts it in SC.
It’s not the game tallying these actions – it’s python scripts written by researchers.
Ladder replays do not have any mouse movement in it. They do not save your mouse position, they do not save its movements. They do not save your camera binds. So:
You can’t make a script, that would count something, that is not part of the replays.
What happens when a unit is commanded to move to a certain location? Ah yes that means the mouse was at that location. Hmm. How about that.