Beruflich Dokumente
Kultur Dokumente
Control
Dan Simon
Cleveland State University
1
Neural Control Architectures
1. Inverse model approach
2. Direct control (derivative-free training)
3. Reference control learning
4. Direct model reference adaptive control
5. Indirect model reference adaptive control
6. Fixed stabilizing control
2
Inverse Model Approach
Two-step approach:
1.Train a neural dynamic system model
2.Train an inverse model to be the controller
Syste
First we train a m +
Inpu Error
neural network t
to model the Neural
dynamic Model
system. Step 1 Learni
ng 3
Inverse Model Approach
+
Referenc Neural Neur Error
e Controller al
Mode
Step 2
l
Learni
Two neural nets in series one neural ng
network
Backprop: Compute derivative of error w/r to controller weights
Backprop: Change controller weights to minimize tracking error
Although we backpropagate derivatives through the neural
model, we do not modify the weights of the neural model
4
Direct Control
Learnin
g
Referenc + e Neural
e Syste
Controll m
er
5
Reference Control Learning
Learning The ANN learns to
mimic the optimal
Neural + controller. Then the
Controll ANN can replace the
er optimal controller.
Ref e
Optimal Syste
. + Controll m
er
a , b
System structure is Learning
given. y +
ANN estimates ANN Model
parameters. and System
Parameters used in ID
controller.
Ref e Standar u y
. + Syste
d m
Controll
er
8
Indirect Model Reference Adaptive
Control
ANN Model and System ID:
b1
uk1 yk
bm
u k m y k +
a1 +
yk1
an
ykn
Weight adjustment
9
Fixed Stabilizing Control
Inverse
Model
Learnin
Referenc + e Standar g
e + Syste
d m
Controll
The standard controllerer
stabilizes the system
The ANN (inverse model) adjusts its weights until the standard
controller output is zero, which means that tracking error e = 0.
The ANN gradually takes over the control function.
10
References
M. Hagan and H. Demuth, Neural Networks for
Control
K. Astrom and B. Wittenmark, Adaptive Control
W. Zhang, System Identification Based on
Generalized ADALINE Neural Network
B. Kosko, Neural Networks and Fuzzy Systems
11