• Advertisement
Sign in to follow this  

Stupid Ideas + Lack of time (2)

This topic is 3003 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

Hey there The thread was closed (http://www.gamedev.net/community/forums/topic.asp?topic_id=550765) so I thought it would be "ok" to start a new one, I liked the ideas, didn't have time to check our HALF of them or even understand, work's for tomorrow, so I'm kinda screwed I managed to make a "good" stabilizer for the hover, it does keep it "stable" tho not in the same place, it just makes it aligned (with some oscilation) using engine's thrust. I'm having problems with the neural network tho... Here's what I have Simulated hovercraft Stabilization algorithm Functional Neural Network with backpropagation So, I'm training this neural network for sometime now with trial and error (changing momentum, learning rate, number of hidden layers) and progress has been SLOW, my mean squared error is like 0.003 which is NOT ok for my purposes (given 4 outputs) Aside from that problem, if I want a neural network to mimic some problem-solution, what should I have in mind? I'm using the input for my algorith, the output for comaparrison and 3 hidden layers Right now the configuration is like this 8 16 16 16 4 tho that's totally random =/ How can I be sure that my neural network is "big enough" to mimic my algorith? Thanks again!

Share this post


Link to post
Share on other sites
Advertisement
I am not sure how you train your network. How do you know if a particular output is "correct"?

One interesting way to use a NN for this type of situation is as an approximation to the value function. This approximation can be refined using the Bellman equation. That can be converted into some mechanism for training). Has anybody tried something like this? I can flesh out a bit what I mean if anyone wants me to.

Share this post


Link to post
Share on other sites
It's a 3D simulation, I can see what's going on, and as I said there's a stabilizing algorithm (that works) training the network, the results may not be the best, but the do stabilize the hovercraft, so, they're correct, and that's the desired output for the neural network

It "worked" but not with the inputs I wanted (I took a more "procecessed input" as entry) now I'm trayning with the same inputs as the deterministic stabilization algorithm

Took me 3.5 hours to train with this pre-processed input, I'll leave it on overnight to see what happens with the new inputs. Note that it's a physx simulation... so the training is in real time (the hover has to actually takeoff and fly a little over and over and over in various configurations).

Share this post


Link to post
Share on other sites
Quote:
Original post by alvaro
I am not sure how you train your network. How do you know if a particular output is "correct"?

One interesting way to use a NN for this type of situation is as an approximation to the value function. This approximation can be refined using the Bellman equation. That can be converted into some mechanism for training). Has anybody tried something like this? I can flesh out a bit what I mean if anyone wants me to.


This is reinforcement learning (Q-learning) with ANNs; if you google these terms together you'll find many papers. :-)

Me, I'd prefer another function approximator (e.g. b-splines), but definitely this is done. Haven't implemented it myself though.

Share this post


Link to post
Share on other sites
I think I have plenty of material to study from now on, I informed myself on what topics should I study before getting into PID controllers, so thanks (ALOT) everyone...

The algorithm worked and I think I managed to pass =p

That's more like something I wanted to do for sometime (not just school work) so I'll probably keep going on this topics.

Thanks again everyone (from this topic and from the 1st one)

Share this post


Link to post
Share on other sites
Sign in to follow this  

  • Advertisement