I use Matlab neural network toolbox for estimating and prediction. In my problem some input parameter values are discrete and qualitative, For example three type of path (A, B, C) as qualitative values and number of pass(1, 2 ,3 ,...) as discrete values.
Is it possible to consider defined number instead of qualitative values and different state?
Is it possible to use directly discrete values in neural network?
From my humble experience:
- no problem with discrete values (including the NN toolbox)
- for qualitative inputs, there is a problem. You cannot code A, B, C as 1, 2 and 3. Because de distance between A and C is then larger than the distance between A and B. You should code A as [1 0 0], B as [0 1 0] et C as [0 0 1] for example.
You can contact me if you need more help!
From my humble experience:
- no problem with discrete values (including the NN toolbox)
- for qualitative inputs, there is a problem. You cannot code A, B, C as 1, 2 and 3. Because de distance between A and C is then larger than the distance between A and B. You should code A as [1 0 0], B as [0 1 0] et C as [0 0 1] for example.
You can contact me if you need more help!
thanks for your response to the question.
Did you mean I should use binary instead of qualitative inputs? (for example using [1 0 0 2 5] instead of [A 2 5], [0 1 0 6 9] instead of [B 6 9] and [0 0 1 3 7] instead of [C 3 7] , so in this example there are five inputs instead of three inputs)
Yes, the same way that an output class is coded in classification.
It is not perfect, but it works.
I would be interested to see other suggestions!
Hi Ali, hi all,
I agree with Amaury in general.
For the quantitative values, you may get into a space covering problem if you use them as such (and by this I mean you should anyway start by "normalizing" them, i.e., arrange for them to be comprised in an interval such as -1;1 or 0;1).
If you run into the space covering problem you could then use a coding that I would call "topological". It is easy to do, here it is what this means.
Let's imagine your values range from 1 to 9 (it's an example, it does not matter much what your actual values are, you need just to adapt my example; it works just fine for continuous values too). For each value you would have an associated pattern of activation over a certain predefined number of input units; let's say 45 units. Imagine this pattern as density values from the Gauss law (i.e., values of the "height" of the normal law curve). The number 5 would be an input pattern of values that are bigger and bigger as you go from input unit 1 to input unit 23, maximum for unit 23, then decreasing again (and symmetrically so) as you go from unit 23 to unit 45. The number 1 would be an input pattern of values such as the highest would be over the first unit and then decreasing as you go from unit 1 to unit 45. The number 9 would be an input pattern of values such as the highest would be over unit 45 and then decreasing as you go from unit 45 to unit 1. I think you've got the idea.
Good luck with your simulations!
Cheers,
SCM
- Badges
- Science topic