I am new to the AI - need suggestions

[Roda Ani]

Hi,

I have as a task to think how to solve a prediction problem using ML/NN. I have the input data which consists of i.e. 5 features which can be combined in different way (can take several values). For each combination a specific output is generated (for example a metric out of simulation). This output has some limitation regarding the max and min value. The idea now is to build NN which is able first to map X->Y F(X),in the second step be able to check in which variations the output is close to min or max value and then propose a xnew VARIATION for a new simulation at which the max or the minimum value could be located.

I would like to have a semi-supervised NN which started with same historical data and then is trained with the new data.

Is it possible to do this with the NN, for example which architecture is more suitable? How can we program neurons to behave as model builder and at the same time predict based on limit values some worst cases?

I hope is not so messy, but for the moments that's all.

[Yoshiyuki Kobayashi]

Do you mean to set upper and lower limits for predicted values?
In the simplest way, you can use Sigmoid function before loss function (e.g. SquaredError) after the last Affine function.
At this time, prepare objective variable in the dataset with 1 for the maximum value and 0 for the minimum value.

[Roda Ani]

Thank you a lot for your answer.

What's is challenging in my problem is how to model a neural network that has:

1. an Xi input which is a vector of features  (e.g. hundreds or thousands of variations are possible)

2. Yi is a scalar value which generated by simulating Xi

Because I am not able to simulate all Xi space I want to use Yi (output of each vector) to give me a path where to look for the new Xinew where the max Yi is located.

First mapping xi->yi (creating a model and updating the function each 5 iteration)

and based on yi (for max or min values ) propose a new vector features Xinew within the space.

It is complicated since I want to make an active learner but I am not sure if Neural Networks (FFNN could be able to perform in such case?). Moreover how to build a model that gives me samples as output.

[Yoshiyuki Kobayashi]

Neural Network Console currently supports only static dataset.
In other words, you can not add data points in the training process.
So, in the case like active learning, it would be better to program in Python using a library like Neural Network Libraries.
https://nnabla.org/

Even when programming with Python, you can still utilize Neural Network Console for designing neural network structure.