the machine will see that most of the people that get reported for “Throwing” are mostly playing certain group of heroes
so it will see something like this:
Playing torbjorn?
Reported for “throwing”
Its a thrower, Ban
Playing Torbjorn/any other hero
Reported for “Throwing”
Hey this doesnt make sense, one of the conditions isnt making sense, this person must not be throwing
What if the “Machine” learns from the wrong source?
what if when the machine learning is released, people go crazy to report, one tricks and off meta players and the machine takes it as a filter “have a x amount of hours on a hero” as a “throwing” filter?
Remember, We have seen this before:
Youtubes desmonetization bots.
the automatic report and ban system.
if you played them in comp, you would realize that its impossible, even if you do good, people will report you… you have to carry at all times non stop, even if you do you get reported anyways and blizzards customer support just tells u “lul no proof”
look at my comp stats, i dont even try to play them in comp in fear of getting reported
I’m not commenting on whether it’s going to work the way you (or any other player) thinks it should. I’m saying that it’s not quite as simple as an ‘if/then’ process and acting like it is is a massive understatement.
Academically and on cutting edge stuff it is. Blizzard probably has better but let’s be honest most places wont go far beyond import sklearn.ensemble.RandomForestClassifier or something equally mundane.
I’ve taken some ML in compsci and honestly it’s only ever as good as the person that designs around the algorithm. The algorithms are detailed and crazy wicked cool but the most critical thing of the success or failure of any sort of ML is the data being put in, the interpretation of output, etc.
Just because someone implements a machine learning algorithm 100% accurately (which tbqh is much easier with libraries like tensorflow floating around) doesn’t mean it’s good. It’s the design considerations around the ML and how it will be used that determine if it’s effective.
they did say that they’re actually developing this machine “excitedly”, which i think translates to “we’re actually working on this thing to make it work, instead of simply throwing a bunch of keywords into a filtering box”.
I don’t think you know how machine learning works in any manner or form from what you have portrayed.
To give you a better idea - for machine learning to work - it actually requires human assistance to tell the machine what is right and what is wrong. It then uses the data gathered to make accurate judgement calls.
For example first it gathers all the reports
“X” player is throwing, because they were AFK.
The Machine Gathers data from the match and attempts to draw a conclusion. A human then tells the machine if the conclusion is right or wrong.
If it detects the word “AFK”
The machine will learn to check for movement patterns of the individual and if that individual stopped moving, and if they did for how long. Did they move their aimer? or just stood still?
Though you may not know it, movement maps are tracking mechanics used by some game developres, such as Counter-Strike go.
This tells developers what is the most common paths players use, how often they use it, and where players die the most often. This will let them adjust maps as necessary in a better fashion.
For example the “Replay” mechanic is not actually “Video recorded” Replay. It is actually the game recreating the entire scene with the data it gathered.
Which means it is tracking all that data, as all that data has to be authenticated by the server anyway.
So essentially machine learning will look for specific keywords, and try to match it with the data. It will use a human to interpert the data for accuracy and then once it hits a very high accuracy rate, it can run more autonomously.
Information it is unsure about can be moved to a CS agent for final approval.