This just proves that Naruto run is the most efficient way to run
This probably would be the more efficient way to run if we didn't get tired from swinging our hands around like that. It's all about the rules the programmers put into the simulation.
I wanna see it do parkour
[media]https://www.youtube.com/watch?v=pgaEE27nsQw[/media]
Yeah but can it control a walking penis and dinosaur while being pelted with boxes.
This was posted by someone on facepunch in another thread ([url]https://facepunch.com/showthread.php?t=1562470[/url]):
[video=youtube;Ul0Gilv5wvY]https://www.youtube.com/watch?v=Ul0Gilv5wvY[/video]
For anyone interested [URL="https://deepmind.com/blog/producing-flexible-behaviours-simulated-environments/"]here's the less excitingly named Deepmind post and papers[/URL]
[quote]We highlight several solution strategies for each terrain in the supplemental video, [B] including less visually appealing ones. [/B] [/quote]
:v:
As cool as this is genetic algorithms and machine learning for this sort of stuff has been around for years and years, and at the end of the day it comes down to a shit load of trial and error until it spits out something that meets the criteria, in this case go from point a to point b.
They just keep trying different combinations of muscle impulses based off the input it is receiving millions of times until something gets them where they want to be. Humans ability to learn things without this sort of non stop, meticulous trial and error is a very different kind of "learning".
Super fascinating stuff to me and I use these concepts in my work so I don't want to trivialize the results, but sometimes these videos and articles summarize this sort of thing as if A) it hasn't been done before and B) make it sound like something it isn't.
[t]http://i.imgur.com/huHF4Q8.png[/t]
[QUOTE=Socram;52463803]They just keep trying different combinations of muscle impulses based off the input it is receiving millions of times until something gets them where they want to be. Humans ability to learn things without this sort of non stop, meticulous trial and error is a very different kind of "learning".[/QUOTE]
How so?
[QUOTE=Drury;52463820]How so?[/QUOTE]
Trial and error is a good way of learning for humans as well, but we also learn behaviors from observing other people.
[QUOTE=01271;52463743][media]https://www.youtube.com/watch?v=pgaEE27nsQw[/media][/QUOTE]
Heh, so that's where that the SWANGIN clip came from.
I'll be super excited when games can effectively use a system like that.
[QUOTE=Drury;52463820]How so?[/QUOTE]
Yes, the notion of "inference" is not something that a computer is capable of in a traditional sense, at least currently. While trial and error is certainly something used by humans to learn, the way we are able to process things we watch others do, or process spatial information (knowing a star shaped block clearly won't fit in a circular hole with out you actually needing to test it), works in ways that computer can't emulate.
Does it really matter at the end of the day? Probably not, especially as processing power gets cheaper and more effective each year. This area (machine learning, generally speaking) is becoming more and more prevalent in the software we use today, not only because of huge developments in the algorithms behind them (which again, I've massively simplified), but also leaps and bounds in the capabilities of CPU's. Computer are REALLY good at trying/processing tons and tons of information sequentially, in ways that humans can't. That's why they "learn" the way they do.
[QUOTE=Laserbeams;52463845]Trial and error is a good way of learning for humans as well, but we also learn behaviors from observing other people.[/QUOTE]
we also have instructions pre-coded in our biological structure
a human raised in isolation (the wild) will still learn how to walk and run like other humans, because it's what's efficient for our body and what's coded into our genes
the base programmer inputs act as substitutes for this, but real life genetic memory is much more stringent
[QUOTE=01271;52463743][media]https://www.youtube.com/watch?v=pgaEE27nsQw[/media]
Yeah but can it control a walking penis and dinosaur while being pelted with boxes.[/QUOTE]
I believe this is a far more elegant solution than the Google algorithm.
[editline]13th July 2017[/editline]
[QUOTE=Zyler;52463766]This was posted by someone on facepunch in another thread ([url]https://facepunch.com/showthread.php?t=1562470[/url]):
[video=youtube;Ul0Gilv5wvY]https://www.youtube.com/watch?v=Ul0Gilv5wvY[/video][/QUOTE]
This has to do with animation blending, not really physics based gait generation from scratch.
Reminds me of a vid someone posted awhile back.
[video=youtube;cRBoto9OMkI]https://www.youtube.com/watch?v=cRBoto9OMkI[/video]
[url]https://facepunch.com/showthread.php?t=1523623&p=50551317[/url]
[QUOTE=Socram;52463867]Yes, the notion of "inference" is not something that a computer is capable of in a traditional sense, at least currently. While trial and error is certainly something used by humans to learn, the way we are able to process things we watch others do, or process spatial information (knowing a star shaped block clearly won't fit in a circular hole with out you actually needing to test it), works in ways that computer can't emulate.[/QUOTE]
But you have to have tested this yourself as a toddler via trial and error one way or another. Even if we extend this scenario to putting a key in a keyhole for the first time, it's a trial and error action where you apply the blocks puzzle knowledge and try and see if it extends to this new problem - you may find that yes, the key does slide in, but the door is still locked. You may try to push the key further to no avail, or - just on a whim - twist it. The door unlocks and you know how to unlock doors with keys now, so you keep that knowledge for later (it may help starting a car some day, which again may require trial and error if you try to steal dad's car for a joyride). Our process of dealing with new situations is vastly built on mixing and matching our previous knowledge (which may or may not be gained from trial and error) combined with trial and error, exactly what the AI does on a smaller scale here.
Due to the way computers work it may not be a 1:1 simulation of how human brains do their thing but the general idea is basically the same. Much like CGI, even if the workings under the hood are vastly different from reality, the end result we perceive is quite familiar, maybe to such a point that it's hard to tell the difference between simulation and the real thing.
There's actually open source simulation software to make hilarious things like this called [url=https://blog.openai.com/roboschool/]Roboschool[/url]. Example:
[vid]https://storage.googleapis.com/joschu-public/demo-race.mp4[/vid]
Tbh imagine how interesting it'd be if we could implement quantum computing and get an AI to play an FPS game effectively
more so than what deepmind has already done in games (like GTA5)
Alternatively create an FPS game based around building guns and teaching an AI how to play cooperatively. And hell, imagine how that can be applied in non-lethal military operations.
[QUOTE=J!NX;52464084]Tbh imagine how interesting it'd be to have an FPS game where you build guns and have an AI companion that learns how to play from you
you basically have to teach an AI how to play the game + build guns by teaching it how to do everything :v:[/QUOTE]
Drivatar AIs in Forza games use a sort of machine learning to drive similar to human players.
[url]https://news.xbox.com/2014/09/30/games-forza-horizon-2-drivatars/[/url]
It's more of a case of "monkey sees monkey does" than "trial and error" but hey.
[QUOTE=Socram;52463867]Yes, the notion of "inference" is not something that a computer is capable of in a traditional sense, at least currently. While trial and error is certainly something used by humans to learn, the way we are able to process things we watch others do, or process spatial information (knowing a star shaped block clearly won't fit in a circular hole with out you actually needing to test it), works in ways that computer can't emulate.
Does it really matter at the end of the day? Probably not, especially as processing power gets cheaper and more effective each year. This area (machine learning, generally speaking) is becoming more and more prevalent in the software we use today, not only because of huge developments in the algorithms behind them (which again, I've massively simplified), but also leaps and bounds in the capabilities of CPU's. Computer are REALLY good at trying/processing tons and tons of information sequentially, in ways that humans can't. That's why they "learn" the way they do.[/QUOTE]
You are far behind the current science
[url]https://deepmind.com/blog/imagine-creating-new-visual-concepts-recombining-familiar-ones/[/url]
[QUOTE=Laserbeams;52463733]This probably would be the more efficient way to run if we didn't get tired from swinging our hands around like that. It's all about the rules the programmers put into the simulation.[/QUOTE]
i wonder if they could put a rule where it has "energy" and each movement depending on how fast and how much it does it drains it, and it has to get to a location without being deprived of it
would it move like us eventually
[QUOTE=Drury;52464101]Drivatar AIs in Forza games use a sort of machine learning to drive similar to human players.
[url]https://news.xbox.com/2014/09/30/games-forza-horizon-2-drivatars/[/url]
It's more of a case of "monkey sees monkey does" than "trial and error" but hey.[/QUOTE]
Now I wonder how this could be applied to shooters in the future. once it gets more complex and a mix of trial and error + monkey see.
imagine UnrealTournaments bot nodes, except with the 'memory' of many players paths, including very complicated ones + rocket jumping being accounted for, as well as hot maps of the most popular spots in the game.
If you wanted to get REALLY complicated, have the AI try and find the least active area that can see the most active area from a distance, and then exploit that with a sniper rifle.
though it'd probably be way more interesting if it was applied to a building game. Learn from humans how to build a contraption and then do so based on whatever variables exist
EDIT:
I just realized that deepmind could theoretically be used for people who lost their legs / were born without legs / had a stroke in order to help them walk.
You could have a computer chip that learns how the user walks and does things and work with them in order to assist them in walking
[QUOTE=J!NX;52464084]Tbh imagine how interesting it'd be if we could implement quantum computing and get an AI to play an FPS game effectively
more so than what deepmind has already done in games (like GTA5)
Alternatively create an FPS game based around building guns and teaching an AI how to play cooperatively. And hell, imagine how that can be applied in non-lethal military operations.[/QUOTE]
I mean there are already fully automated rage hacks for CSGO which people use to level lots of accounts at once. They just walk around the map using the normal bot navmeshes.
Soon: "Google's Deepmind AI teaches itself how to shoot a gun" :v:
[QUOTE=J!NX;52464084]
Alternatively create an FPS game based around building guns and teaching an AI how to play cooperatively. And hell, imagine how that can be applied in non-lethal military operations.[/QUOTE]
Not exactly the same, but I did work on a defense R&D project that involved AI and military logistics/planning. Basically, the defense industry is pretty up to date with everything that's going on.
[QUOTE=SirJon;52464211]Soon: "Google's Deepmind AI teaches itself how to shoot a gun" :v:[/QUOTE]
"Deepmind learns how to manufacture simple machines"
"[I]Deepmind learns how to secure a building[/I]"
"[I][B]Deepmind locked us out of the building[/B][/I]"
"[I][B][U]DEEPMIND SEND A CYBORG TO ANSWER THE DOOR[/U][/B][/I]"
[QUOTE=abananapeel;52463989]Reminds me of a vid someone posted awhile back.
[video=youtube;cRBoto9OMkI]https://www.youtube.com/watch?v=cRBoto9OMkI[/video]
[url]https://facepunch.com/showthread.php?t=1523623&p=50551317[/url][/QUOTE]
Is this how Zygote was made? Fuckton of hands.
It looks so happy and free.
Sorry, you need to Log In to post a reply to this thread.