I''ve been doing some NN reseaches, made my small library several examples e.t.c. But what I''ve noticed is that it''s necessary to do all the education at one time. I mean, if later another example should be taught, NN will learn it, but will forget about everything taught before. How should I fight with it? Also, does someone has a good link on reinforcement learning with some examples e.t.c.

Normal NN do not forget what they have learned. May be you are talking about Hopfield networks?

I''m guessing you are talking about feed-forward networks and that you have been training them using a gradient descent algorithm like backprop...

Once a network has been trained, if there is subsequent new data that it must learn to classify, then you must retrain the network using the old data set + the new data.

You can find an online book about reinforcement learning here:

http://www-anw.cs.umass.edu/~rich/book/the-book.html



ai-junkie.com

you could also fully train a network with given data, fix the topology and the weights then. Now if you get new pattern, you could do something cascade-correlation like.

but I guess reinforcement learning is here a good solution, although the net might still ''delearn'' some stuff using reinforcement learning.

I see the problem. Ideally you want to train the network on the old data + new data, but you don''t want to keep around the old data. So here''s a wacky idea and you can tell me if it sounds too stupid to work:

* Make a copy of the NN in question, and use it to reproduce the old data. *

It is a big vague, but then the specifics would depend on the NN in question. I think it''s naive to use one algorithm for all possible implementations.

Tom

Re-inforcement learning is what I''ve been thinking about, but how does it relate to NNs? I''ve been reading some reinf. stuff, but they were related to problems, where rules or behavior-patterns were being designed, approved/punsihed, but how to incorporate it in Neural Networks? I thought about keeping old data, but it would be too memory and CPU intense to retrain nn with the whole bunch again and again, save all the training examples.
And what is cascade-correlation?

As for using a copy to teach, that wouldn''t work right, as by that teaching it will overwrite new examples and in the end there will be the same old NN.

It sounds like what you need to implement is sequential learning rather than batch learning. There has been a fair bit done on learning the parameters of ANNs while utilising the network to classify data (i.e., training during use). I don''t have my bibliography on this computer so I cannot give you direct references at this time. You should be able to find plenty online regarding the topic though.

Cheers,

Timkin

I tried searching for sequential learning as I feel it''s the right way to do it, but was getting only some buggy servers having links, not files, some required passwords and some didn''t have anything.
Could you list a few links?

And a second question. How does "a cycle" in a network with non-linear struct works? In linear, it''s just w1*i1+w2*i2+...+wn*in
, but how does it work in networks that might have any type of connection possible?