I am working on an ITS technique which is called dynamic lane grouping at signalized intersections in which the lane group will be chosen based on the movement traffic demand. The inputs of the neural network model will be 12 volumes ( 3 movements (left, through and right) at 4 approaches) and I have 4 outputs (the optimum lane group at each approach) which are represented by integer values (from 1 to 10). Should I convert the inputs and outputs to any format before applying neural networks?
Hi,
Yes, you must normalize values between [-1 to 1] for a better Learning when using neural network approach.
The only difference that matters is whether one uses continuous or discrete values, since they affect the capacity.
Hi,
Yeah if you normalized the input values between -1 to 1 or 0 to 1 it will be easy and efficient for learning and for the output, based on the activation function you are using in the output layer normalize your expected outputs. Eg: if you use the sigmoid then your output will be between 0 to 1 and for tanh -1 to 1.
Normalizing input values between 0 to 1 will be helpful for suitable learning of neural networks.
Hi,
in case you appreciate further information, there is a great FAQ hosted by SAS. See ftp://ftp.sas.com/pub/neural/FAQ.html, and for the specific question of scaling inputs: ftp://ftp.sas.com/pub/neural/FAQ2.html#A_std_in
Hope that helps
Normally, double format is accepted by any classification based network. So, you can convert all the images into double and try.
you can normalize inputs using formula: x_norm = ((x-x_min)(d2-d1)/(x_max - x_min))+d1. xmin, xmax - inputs interval. d1,d2 - normalize interval. i know recommends than d1, d2 are the values as output interval of neuron activation function.
As some others already mentioned, it is improving the network performance if you normalize your dataset before training your network, typically to a 80% of the value projectd to the intervall [-1,1].
In many of the applications, I have normalized input values in the range -1 to 1, it has nicely worked. Further, it would avoid the product ( weight ×input) reaching overflow.
I think you should convert it. You should configure input, output and target data. You can configure optimum input and output and give a nominal weight based on its rank (the weight is different with the weight between layer to layer in neural network) . These configuration act as a prior data and the weight can be used as target.
If the data is represented by integer values then it is fine to work on same. Otherwise it also depends on the type of utility used for implementing ANN. Like in matlab or in r programming integer values are pretty much fine.
no I think it is not necessary to convert them, you just have to introduce them in digital format
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence