Post by tkorrovi on Dec 2, 2004 19:55:45 GMT -5
Project page:
sourceforge.net/projects/adsproject
There are no known bugs. But the biggest problem is that such system develops very slowly. It took several months to train the test structure, di.str. Now it's trained to play the game of Nim, training would certainly last much longer, as it's much more complicated, but the problem is that nobody knows how much longer. And because of that, when it does not at present do what it supposed to, it doesn't mean that it isn't able to. The latest structure is under the project's home page ( sourceforge.net/projects/adsproject ), I shall run it, and upload it from time to time. But still, this is a big problem, because of low speed it cannot be tested positive or negative.
It may seem that because of this, the project is not successful. But it is successful, the test structure di.str is trained to write certain small word, when it predicts that the same word would be inputted, and it tries to find another alteratives when the input is different. You can see it when you download the program. Yes, I understand that this is very little what it does, but considering a low speed of learning, it is all what is possible to achieve at present. Consider that such behaviour was trained to the system, which consists of nothing more than a few lines connected together, with no knowledge more than that. And it is because of that, at least in theory, capable to model any process, being unrestricted, as no neural network or likely no other system is. Considering the importance of that, it's worth to study such or similar systems. I hope that I said everything, which is relevant.
sourceforge.net/projects/adsproject
There are no known bugs. But the biggest problem is that such system develops very slowly. It took several months to train the test structure, di.str. Now it's trained to play the game of Nim, training would certainly last much longer, as it's much more complicated, but the problem is that nobody knows how much longer. And because of that, when it does not at present do what it supposed to, it doesn't mean that it isn't able to. The latest structure is under the project's home page ( sourceforge.net/projects/adsproject ), I shall run it, and upload it from time to time. But still, this is a big problem, because of low speed it cannot be tested positive or negative.
It may seem that because of this, the project is not successful. But it is successful, the test structure di.str is trained to write certain small word, when it predicts that the same word would be inputted, and it tries to find another alteratives when the input is different. You can see it when you download the program. Yes, I understand that this is very little what it does, but considering a low speed of learning, it is all what is possible to achieve at present. Consider that such behaviour was trained to the system, which consists of nothing more than a few lines connected together, with no knowledge more than that. And it is because of that, at least in theory, capable to model any process, being unrestricted, as no neural network or likely no other system is. Considering the importance of that, it's worth to study such or similar systems. I hope that I said everything, which is relevant.