Hi everybody! I played around a bit with the Support Vector Machines, which are supposed to perform pattern recognition (both spatial and temporal) at least as good as the Neural Networks and have few or no problems with the training. I'm not a game developer and so I'd like to team up with someone who knows the business to give'em a try and see what comes out. Anybody interested? -JustADude

What comes out from what? If you have a goal, please share it ;)

Hi Nyarlath,

I think it would be interesting to use the SVM in a simple video game and see if they (the SVM) can be used to learn and possibly forecast spatio-temporal patterns (like, for example, player movements). Virtually anything one can do with neural networks can be done with the SVM, with the advantage that the SVM are easier to train and perform usually a bit better.

Fell free to contact me hpe8hud1busc72u@jetable.com
we could chat about it in a cafè in Zürich downtown...

best
-dude

Quote: Original post by Nyarlath
What comes out from what? If you have a goal, please share it ;)

Quote: Original post by justadude
Hi Nyarlath,

I think it would be interesting to use the SVM in a simple video game and see if they (the SVM) can be used to learn and possibly forecast spatio-temporal patterns (like, for example, player movements). Virtually anything one can do with neural networks can be done with the SVM, with the advantage that the SVM are easier to train and perform usually a bit better.

Hmm, predicting player movements might be tricky with SVMs. At least the ways to do that that immediately come to mind.

You are correct that SVMs outperform neural networks on many tasks, the best examples of which are handwriting recognition and related stuff. And SVMs don't have problems with local minima which plague neural networks. However, on complex, structured data - like player movements, I am guessing - both simple neural networks and SVMs might not do so well. More complicated neural networks might actually do better (convolution networks, for example).

What you might do is extract a few 'features' from the player movement data, and as a final step train an SVM on that. That would work, but deciding on the features would be the hardest part, and SVMs don't help with that (picking the features is pretty much deciding on the kernel function for the SVM, and that is a parameter to the SVM, not something it finds on itself).

I think the most fundamental questions you need to answer for either NN or SVM training are:

1) Where will I get my data? You're going to need training instances to provide to the learner and in an FPS type of environment, I'm not convinced you can amass the necessary number of scenarios to ensure adequate learning. Of course, you could use reinforcement learning but that might take a while.

2) How will I represent the environment for input to the learner? Representing the environment as a grid with flags in certain locations describing obstacles, enemies, etc may work, but you'll get stuck with a learner memorizing a particular environment.

And, I don't know if I agree about the ease of training an SVM relative to that of a NN. I can program my own congugate gradient minimizer to get a very good NN fast, but I don't think I'm capable of writing the quadratic programming solution for an SVM. Or maybe I just don't know enough about optimization theory yet. 8^)

Quote: Original post by kirkd
And, I don't know if I agree about the ease of training an SVM relative to that of a NN. I can program my own congugate gradient minimizer to get a very good NN fast, but I don't think I'm capable of writing the quadratic programming solution for an SVM. Or maybe I just don't know enough about optimization theory yet. 8^)

Well, I wouldn't write either from scratch - there is high-quality code already available, e.g., libSVM for SVMs.

Aside from ease of programming, there's another issue - ease of training. NNs have local minima, so training them can be tricky at times. SVMs don't have that problem.

hey ne body has some more information on kernel functions.. plz share it

@ashish1512

I can help you with the kernel functions if you give me more information about which is your problem and if you had studied the feature space, if so please post some information about it.

If you're planning on doing real-time learning, SVM is a little bit too much in my opinion.