My thoughts on AI. (It's not about games in particular)
Confusing text removed. I hope that's ok in this forum.
But if you really want to see what was confusing by reading the text yourself, click the 'hstory' button on the button of the post to see the deleted text.
The current thread related text I want to talk about has 14 steps and is a few posts down.
But if you really want to see what was confusing by reading the text yourself, click the 'hstory' button on the button of the post to see the deleted text.
The current thread related text I want to talk about has 14 steps and is a few posts down.
For the music piece - why specifically ANNs and not, say, Markov chains?
I suspect you'd get much better music out of a Markov model or something similar rather than a neural network.
(Snark for Dave's sake: you could probably get equally good results from a random tone generator with a lot less effort. I jest, but ANNs aren't exactly the right tool for this kind of job.)
It seems to me like you're taking ANNs and trying to hit everything with that particular hammer, instead of looking for much more suitable modeling techniques and approaches to simulation and problem solving.
I suspect you'd get much better music out of a Markov model or something similar rather than a neural network.
(Snark for Dave's sake: you could probably get equally good results from a random tone generator with a lot less effort. I jest, but ANNs aren't exactly the right tool for this kind of job.)
It seems to me like you're taking ANNs and trying to hit everything with that particular hammer, instead of looking for much more suitable modeling techniques and approaches to simulation and problem solving.
Speaking as an AI programmer and an ex- music theory and composition major, this is seriously twisted. Apoch is correct. All a NN is good at doing is detecting patterns and spitting out some sort of assumption. There has been automatic composition software designed quite successfully based on Markov chains as he suggested.
I'm really at a loss for what it is you are attempting to accomplish here. Your posts do tend to wander a little bit.
All a NN is good at doing is detecting patterns and spitting out some sort of assumption.
Google "deep belief net" or "Geoffrey Hinton". You can get a NN to do the same things a Markov chain would by running the network backwards.
Jeremy; the idea of treating data as some form of geometry is at the core of most of these sorts of algorithms, so you've got the right idea.
A neural network treats the input as a coordinate-- we use the term Vector instead of coordinate, but I'll use coordinate here for simplicity. If you have two inputs, you have a two dimensional coordinate. If you have 50 inputs, then you have a 50 dimensional coordinate.
The network, through it's weights and activations, will "project" the input coordinate on to a lower dimension. Its like how your shadow is a 2D projection of your 3D self. If you're network has only two outputs, then you can take the 50 dimensional input and use the network output to draw its location in two dimensions.
A network that performs well will make it so that inputs that are "similar" end up being close together when they are projected on to a lower dimension. If your input data describe fruit, and had 10 input parameters, and the network had two outputs which you treated as x and y coordinates on a graph, you would expect all apples to be clustered together, all bananas together but far from apples, etc..
Definitely look up Markov chains. There was a post on gamedev in the AI forums a few months ago where someone shared some results of using Markov chains to generate music.
A neural network treats the input as a coordinate-- we use the term Vector instead of coordinate, but I'll use coordinate here for simplicity. If you have two inputs, you have a two dimensional coordinate. If you have 50 inputs, then you have a 50 dimensional coordinate.
The network, through it's weights and activations, will "project" the input coordinate on to a lower dimension. Its like how your shadow is a 2D projection of your 3D self. If you're network has only two outputs, then you can take the 50 dimensional input and use the network output to draw its location in two dimensions.
A network that performs well will make it so that inputs that are "similar" end up being close together when they are projected on to a lower dimension. If your input data describe fruit, and had 10 input parameters, and the network had two outputs which you treated as x and y coordinates on a graph, you would expect all apples to be clustered together, all bananas together but far from apples, etc..
Definitely look up Markov chains. There was a post on gamedev in the AI forums a few months ago where someone shared some results of using Markov chains to generate music.
Every time you post something I try to read it, but after just a couple of sentences, some "formula" appears with no explanation whatsoever as to what anything in the formula means. I feel you didn't give me a chance to understand what you are saying and I quickly lose interest.
If there is some merit to your ideas, I can't figure out what it is, because your descriptions are not intelligible. I can assure you nobody understood what you meant by "(a=a)" being the input.
If there is some merit to your ideas, I can't figure out what it is, because your descriptions are not intelligible. I can assure you nobody understood what you meant by "(a=a)" being the input.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement