Neural Network based controller

[date.chinmay]

I'm working on a project which deals with temperature control of a room. the idea is to control temperature within a limit.

I have prior data which i used to train a neural network which is 99.5% accurate. there are 5 inputs (x , y , z , A , B) and there is one output (T).

Now I want to use this network in a control architecture of some kind. Any suggestions? which should I use?

P.S. The system is highly complex, non linear and dynamic in nature. I have already tried standard inversion of system but it doesn't work as outputs are lesser than inputs.
I'm doing this in Matlab.

 

[Floid]

I am not quite following your question. Typically you would train the network using your inputs and the desired output for those inputs. In essence what you would be doing is trained the NN to become a feedback controller for the 5 inputs.

To use the NN you would then put in a real time system with some types of sensors feeding your x,y,z,A,B values and your T value going to an actuator that effects some output.

 

[date.chinmay]

I am not quite following your question. Typically you would train the network using your inputs and the desired output for those inputs. In essence what you would be doing is trained the NN to become a feedback controller for the 5 inputs.

To use the NN you would then put in a real time system with some types of sensors feeding your x,y,z,A,B values and your T value going to an actuator that effects some output.

Thanks for replying!
let me come straight..
X-Y-Z are my coordinates... A is temperature controlling device.. and B is the Heat dissipated at that point...

All these affect the output i.e. T at that point.

when i say i am trying to invert it... i mean.. I'm trying to make the ANN act as if.. when it reads the temperature in an actual REAL system.. it should predict the input parameters... so as to push it towards a desired state.. Can I email or private message you for more clarification?

 

[Floid]

Sure, you can send me a PM but I am not sure I can be much help. It sounds like you want you NN to provide the feedback for a controller so that based on a temperature T you supply the controller with input (x, y, z, A, B) that go to an actuator that is capable of driving T one way or the other.

Given that, what type of learning do you plan on implementing?

If you are doing online (reinforcement learning) then is your environment dynamic while you are trying to train the NN? If so it is going to be next to impossible to train it. The NN may be giving a good output but the environment is natural changing T in the other direction outside of your control making the learning process think you are giving it bad inputs. Or vice versa... you may be giving bad control outputs but the environment happens to be changing in your favor so the NN thinks it is doing good.