rin   yout  error  1
           |     |      |    |
           |     |      |    |
           ↓     ↓      ↓    ↓
          ┌───────────────────┐ 
          ∣     BP Neural      ∣
          ∣     Network        ∣
          └───────────────────┘
             |      |      |
             |kp    |ki    |kd
             ↓      ↓      ↓
            ┌───────────────┐      ┌───────┐ 
rin +    e  ∣                ∣   u  |       |   yout 
────→( )───→∣        PID     ∣─────→| Plant |────┬→
    - ↑     ∣                ∣      |       |    |
      |     └───────────────┘      └───────┘    |
      |                                         |
      └─────────────────────────────────────────┘

Above is the controller I have designed. Could anybody tell me how to train this BP network with the MATLAB NN toolbox? Well, the difficulty lie in that there is no targets for the outputs of the NN. The plant output "yout" is the network input in the next loop. *Edited by moderator to fix ascii graphics* [Edited by - Timkin on November 24, 2005 6:31:08 PM]

If you surround your ascii art with the code and /code tags (embed them in square brackets), then it will preserve your layout. Could you please adjust it so that we can see your architecture correctly.

From what I can make of the diagram, you appear to have drawn a disturbance rejection architecture... and you're trying to generate an adaptive PID control architecture. You could probably obtain a weight gradient via a lagged state error difference, although I suspect this would have some problems.

Interestingly enough, I'm working on a very similar problem in my current research, although we don't limit ourselves to PID regulators (but rather use RBF networks, spline networks, CANFIS, etc). Mostly I'm looking at appropriate architectures for training of such networks on arbitrary plants. I'm fairly sure I have a paper on my desk at work on training PIDs in adaptive networks... I'll check tomorrow and post the details if I can find it.

Cheers,

Timkin

I think you need to put a delay from the feedback, maybe. Otherwise I do not understand how you could realize this.