Creating simple Multi-layer Perceptron in Matlab

I want to train a MLP that gets 3 inputs and generate 6 outputs, how can I build, train and test such MLP?

Suppose I have a robot, whose hand has a (x, y, z) position, but this robot has 6 joints, and for each (xi, yi, zi) the hand is in, there will be a (j1, j2, j3, j4, j5, j6) representing the values of the joints that take the hand to (xi, yi, zi).

I want to build an MLP that approximate this mapping between world-point and joints, and I already collected 1000 pair of (xi,, yi, zi) and (j1, j2, j3, j4, j5, j6) to train the MLP and to test.

So, finally, how can I build, train and test such MLP?

Answer 1

[Herve POSTEC]

Hello

I'm not a big specialist about it but I tried to do the same computation as you. I just simulated a 2 DOF arm (2 angles)

I proceeded the following way : - Create couples of angle values (a1, a2). Save it in variable 0ut - Compute (x, y) position related to the angles (simple trigonometry). Save it in variable In - Then learn a neural net in order to get Out function of In

Results were very bad. I think it's because for one (x, y), there are several possible (a1, a2) and the net doesn't "know" how to distinguish them

The process seemed rather simple but I'm a beginner on machine learning and I didn't succeed.

Hope this helps you a little

Bye