cool ai

[maxxor]

 

[James Kerman]

Welcome to the forum @maxxor.

I've moved your topic from KSP 2 discussion to science and spaceflight.

[maxxor]

no

[maxxor]

this should be in ksp2 the devs can use neural learning to train the game to expect physics calculations and boost performance with 1000fps

[James Kerman]

9 minutes ago, maxxor said:

this should be in ksp2 the devs can use neural learning to train the game to expect physics calculations and boost performance with 1000fps

Oh, I'm sorry, I misunderstood. Moved back to KSP 2 Discussion with my apologies. 

[sgt_flyer]

Interesting video - it could likely help for a lot for normal launches of monster rockets, but would likely need a way to switch to higher quality physics (up to full simulation) as you approach edge cases for the loaded learned set.

as seen in the video, edge cases (fast moving objects) are not registered by the trained AI currently - and ksp can quickly have high speed collisions (staging too early side boosters, etc)

Now, the AI would be needed to be thoroughly trained with rigid bodies, and more importantly to be able to work on random rocket builds 

 

 

[steuben]

Marvelous performance increases and cost reduction thanks to <AI/blockchain/quantum computing/internet/magnets delete as appropriate.> Call now operators are standing by.

Two to three orders of magnitude decrease in processing time... It would be a time of wonders if it pans out.

Training an AI starts at pretty finicky and gets worse. Reminds me of the time they successfully trained an AI for finding camouflaged tanks to tell you if the day was cloudy or not.  And the dataset for stock/DLC KSP would be a bit large. Aside from the sanely, well sane for KSP players, designed craft, you would need a non-trival number of the game engine allows this to fly designs, several completely random designs, and few designs that only look normal under various eldritch geometries. 

And it would probably need a complete retraining anytime a part has its parameters changed.

And then there are the modded parts...

[mattinoz]

2 hours ago, sgt_flyer said:

Interesting video - it could likely help for a lot for normal launches of monster rockets, but would likely need a way to switch to higher quality physics (up to full simulation) as you approach edge cases for the loaded learned set.

as seen in the video, edge cases (fast moving objects) are not registered by the trained AI currently - and ksp can quickly have high speed collisions (staging too early side boosters, etc)

Now, the AI would be needed to be thoroughly trained with rigid bodies, and more importantly to be able to work on random rocket builds 

 

 

Wasn't fast moving objects just an edge case for the model as trained, a model trained for KSP would handle most KSP situations. Also if the engine is that fast you could shorten the physics frame is fast moving situations and drop the distance each frame needs to deal with back into scope.

 

[Gargamel]

13 hours ago, maxxor said:

this should be in ksp2 the devs can use neural learning to train the game to expect physics calculations and boost performance with 1000fps

Perhaps in the future, you could post some context to the random video you are linking, it would greatly help the discussion move forward.   Thanks!

[magnemoe]

14 hours ago, steuben said:

Marvelous performance increases and cost reduction thanks to <AI/blockchain/quantum computing/internet/magnets delete as appropriate.> Call now operators are standing by.

Two to three orders of magnitude decrease in processing time... It would be a time of wonders if it pans out.

Training an AI starts at pretty finicky and gets worse. Reminds me of the time they successfully trained an AI for finding camouflaged tanks to tell you if the day was cloudy or not.  And the dataset for stock/DLC KSP would be a bit large. Aside from the sanely, well sane for KSP players, designed craft, you would need a non-trival number of the game engine allows this to fly designs, several completely random designs, and few designs that only look normal under various eldritch geometries. 

And it would probably need a complete retraining anytime a part has its parameters changed.

And then there are the modded parts...

This looks very good for cloth physic in games, here you have an set number of items like skirts, robes, dresses, trench coats, capes, and long hair you want cloth physic on. However you can have lots of copies of them with lots of npc or players so baking in the limited numbers of styles makes lots of sense. 
However in KSP each construction is unique and it changes with staging and docking 

[maxxor]

you guys arent thinking like an ai or a computer you just see different label cloth isnt the same as rocket. but an ai just handels collision physics its pretty universal if you set the same standards of interaction. almost all parts will bump similarly against eachother. doesnt matter if you have a supercustom cool rocket the parts are still interconnected in a similar way. even if your meme rocket is unique to your ego neural processing will make any game handle a 700 part much easier than if it never existed.
most times of physics and slowdowns are during launch explosions and landings the only connections that are really in effect are seperators, topcones of air, landinglegs. it could probably learn how to expect things to bump around at 4x speed too.

[Rejected Spawn]

I'm just going to chip in and say that this AI thing is really shady and makes wild claims, on top of it being completely unsuitable for systems with too many types of objects all at once. That video claimed a simple simulation of a scrap of cloth was running at TWO frames per second by simply properly calculating it, I find that very difficult to believe if the physics engine used had been even remotely well programmed. The claim that the AI ran it at thousands of frames per second is less difficult to believe - because that simulation is extremely simple and the AI is likely just regurgitating the closest known approximation with terrible accuracy. Not once did the AI presented actually have to work on more than an abysmal number of types of objects at the same time, this is a huge red flag if you want to "sell" an AI that's competent and can be applied to complex systems. The claim that it "only" uses 9MB (or whatever it was) is for a pathetic number of types of objects in play. I'm now going to use math so if you can't stomach that, it's time to stop reading:

To properly calculate the outcome of an impact between two objects you need to know their exact physical shape, weight, speed, rotational speed, angle of impact and point of impact. When you let the same two objects collide thousands of times you can get a number of good examples as well as edge cases - then all you need to do is match the closest one to what's about to happen next time and you have an ALMOST GOOD ENOUGH really fast prediction. Now let's add just one more object, Object3, and run the simulation for a few thousand times between that and Object1 and then again a few thousand times for Object3 colliding with Object2, your database now grew to three times its first size by simply adding A SINGLE OBJECT, this results in slowdowns since the AI has to dig through the much larger database for a match. Now let's ATTACH Object1 to Object2 and collide that with Object3, this introduces yet another factor for the AI to consider: how much force it takes to separate Object1 from Object2. The database is growing. Now let's attach Object1 to Object2 in a slightly differrent spot and we have yet another scenario that causes database growth. Now let's attach Object3 to Object2 and collide with Object1 that is attached to a second copy of Object1. At this point the AI is already using a huge database that is nothing like what that shady video would have you believe, and we're still on THREE object types. How many pieces do you build your rockets out of? I certainly never build a rocket out of just two pieces. I'm not sure at what point the database hits the terrabyte mark but for something like KSP2 I'm pretty sure the database size would make you pale in the face - and it still wouldn't be anywhere close to accurate... not to mention its selling point is completely lost because it would no longer be even remotely fast.

I could go on but anyone who still doesn't understand is not going to listen to reason and logic, I have better things to do with my life. Final nail in the coffin for this topic: KSP2 is - as far as I understand - already developed to the point where it can handle big constructs accurately and it will use some kind of "grouping" of objects that greatly reduces the number of calculations needed in most cases and that still permits simply dividing the group if circumstances demand more finely detailed calculations. This system is close to perfectly accurate, runs fast even on low spec hardware and is already close to done. Anyone who wants to throw out this system 6 months before the game hits the market has no business giving advice to game developers.

[magnemoe]

22 hours ago, maxxor said:

you guys arent thinking like an ai or a computer you just see different label cloth isnt the same as rocket. but an ai just handels collision physics its pretty universal if you set the same standards of interaction. almost all parts will bump similarly against eachother. doesnt matter if you have a supercustom cool rocket the parts are still interconnected in a similar way. even if your meme rocket is unique to your ego neural processing will make any game handle a 700 part much easier than if it never existed.
most times of physics and slowdowns are during launch explosions and landings the only connections that are really in effect are seperators, topcones of air, landinglegs. it could probably learn how to expect things to bump around at 4x speed too.

Yes I'm aware of that. And this might work well for destructive structures too. However the AI has to be trained and the result is then baked into the design. 
This is perfect for clothing as its an limited number of styles, you could treat robes and dresses as skirts , however most rockets tend to be unique and only used once, or the design evolve, they also stage.
I assume the AI creates an simplified model of the clothing where just some points are used for cloth physic. The upper part of an skirt will not be much affected it compared to the hem as an simple example. 

[c4ooo]

On 10/30/2019 at 8:01 AM, maxxor said:

this should be in ksp2 the devs can use neural learning to train the game to expect physics calculations and boost performance with 1000fps

 

23 hours ago, maxxor said:

you guys arent thinking like an ai or a computer you just see different label cloth isnt the same as rocket. but an ai just handels collision physics its pretty universal if you set the same standards of interaction. almost all parts will bump similarly against eachother. doesnt matter if you have a supercustom cool rocket the parts are still interconnected in a similar way. even if your meme rocket is unique to your ego neural processing will make any game handle a 700 part much easier than if it never existed.
most times of physics and slowdowns are during launch explosions and landings the only connections that are really in effect are seperators, topcones of air, landinglegs. it could probably learn how to expect things to bump around at 4x speed too.

wut

The video demonstrates the use of the NN for simulating cloth. There is absolutely nothing to suggest that it would work well for rockets, and the claim that using NN will magically "boost performance with 1000fps" is ridiculous. NN aren't a magic solution to every single problem as some people like to paint them as. 

NN = Neural Networks