Jump to content

  • Log In with Google      Sign In   
  • Create Account

CulDeVu

Member Since 07 Nov 2010
Online Last Active Today, 11:16 PM

Posts I've Made

In Topic: Backpropagation XOR issue

13 December 2014 - 04:25 AM

There's a bunch of reasons why an arbitrary feed-forward neural network might fail to converge to the correct solution. It might converge to some other local minima, or some other issue.

 

If it's just a neural network for XOR that's doing that, then you have an issue. You might want to check your initial weights and biases and make sure they aren't too close to zero, as that can sometimes be really hard to recover from. Other than that, you might have some implementation bug...


In Topic: Working on Graphics engine for Nintendo Systems

12 December 2014 - 01:55 PM

First off, please don't double post, especially the same thing under different names: http://www.gamedev.net/topic/663682-global-illumination-and-shader-engine/

 

Second, you should consider doing a write-up instead of posting a youtube video with a code-dump inside of it. It'd be a lot more accessible that way.

 

I'm not actually affiliated with Nintendo or its dev-kit, so I can't tell you one way or another it'd work, sorry.


In Topic: Backpropagation XOR issue

12 December 2014 - 08:37 AM

When you strip away all of the inter-node fun, backpropagation is just a really fancy version of gradient descent. And like all gradient descent algorithms, it's extremely sensitive to whatever convergence rate you give it. That's why these sorts of things are usually described mathematically as differential quantities, because that's really the only time they'll always reliably work :P

 

So yes, it's possible to get it stuck forever overshooting if you give it a learning rate high enough. 


In Topic: Slope physics

05 December 2014 - 09:58 AM

No, it looks like you're mathing everything okay.

 

http://hypertextbook.com/facts/2007/TabraizRasul.shtml

 

If you don't want your character to move as far, just increase the friction coefficient. The coefficient of wet snow is generally 0.3-0.6 when in contact with rubber, so it it's probably a little higher than 0.2 for snow in contact with a snowboard.

 

Also keep in mind there's a lot of other forces at play here besides just friction. There's additional soft-body collisions happening when the snowboard cuts through the snow, there's the air resistance that goes into affect when the snowboarder goes fast enough, there's the difference between the top layer of snow that's wet and the lower layers that're probably dry that have different affects on the board. In the end, though, your game is going to be an estimate of reality, so don't be afraid to play with some numbers to make things look and feel right. It may not be mathematically correct, but a lot of times, it's all you can do :)

 

I'd recommend dialing up the friction coefficient, and adding air resistance, and that should be good enough for a semi-accurate simulation.


In Topic: Multiple lights give white

31 October 2014 - 11:22 AM

Long story short, you need to tonemap the values in your image. This is because your screen operates under the range [0, 1] and real light operates under [0, inf). This makes sense, because you can keep making a room brighter and brighter by throwing in more and more photons, but your screen can only get but so white. So you need to turn real-world brightness (HDR) into something the screen can display (LDR).

A really easy way to get started is to transform each RGB component in your image by
f' = f / (f + 1)

There are more advanced ways, but that's the jist of it. Throw it into your shader, boom! Your HDR image is now able to be viewed on a screen. I don't think it's able to be done in the fixed function pipeline, someone correct me if I'm wrong.

If you want to know more, start Googling around for HDR rendering.


PARTNERS