Pramod Gupta, Naresh K. Sinha
Mechatronics, No. 10, 2000
Summary
The papers is about the application of neural
networks for “model-free” controllers, so that they can learn online the
composition of the system improving constantly their performance.
Today robotics face the
challenge of flexibility, requiring to operate in real-time, pick up novel
payloads and be expandable to accommodate many joints. Traditional control
methods (direct program control, teaching pendants, inverse kinematics) are not
adaptive and may work well in highly repetitive environments, but not in the case
of uncertainties. Recent studies have demonstrated important characteristics of
learning by neural networks, this methods appears to be adaptive and therefore
ideal for a general robotic system.
·
Model
The papers goal is to
perform an experimental model to get the real-time features and numerical
characteristics of the proposed neuro-control schemes. Basically the neural
network is trained offline to approximate the inverse dynamic model of the
manipulator; the errors during operations can be used to train the neural
network online (through the modified delta-bar rule) The fundamental steps of
the learning algorithm are: weight update
rule, learning rate update and gain update rule.
The weight update rule is:
w1ji(n+1)=w1ji(n)+;ηji(n+1)δlj(n)yi(l-1)(n)+αΔwlji(n-1) where
δlj=γejl(n)fl(wlj(n)),
w1ji(n) is the synaptic weight connecting neuron ith in
the previous layer to the jth neuron in the lth layer, η is the learning rate, which is following Jacob’s
heuristics (please refer to paper).
The learning rate depends
on Dji(n) and Sji(n), the former being the
partial derivative of the error surface with respect to wji(n) and the latter being an exponentially weighted sum of the
current and past derivatives of the error surface with respect to wji(n) and with a constant indicating
the base and iteration number n as
the exponent.
The general procedure for
control systems development involved: control task definition, contro system
hardware setup, modeling parametric identification, control scheme selection
and simulation test, real-time control code programming and implementation and
system testing (the authors analyze only the last 2, since they use a robot,
FLEXROD, where the former steps have already been done.
·
Experiment
The
experiment involved FLEXROD, which has the following characteristics: a
mechanical arm with two DC motors, one encoder per motor, transmission gear and
light weight aluminum structure, a controller amplifier package, a personal
computer interface for communication and programming. The joint velocities are
obtained by differentiation of the join angles, due to noise in the
differentiation a Butterworth filter is used with a cutoff frequency at 100Hz.
Key
Concepts
Controlling by learning,
Neural Networks, Online programming, Intelligent Robots
Key Results
It has been observed that
the qualitative behavior is very close to that obtained in the simulation
phase. The same tasks performed with a PID controller demonstrated results with
a larger error.
Creating control system
based on the neural network producing the largest part of the control input
with a linear compensator generating small compensations is then effective.
Compared with PD controllers it has been demonstrated that there are less
oscillations and a neural network based scheme is then capable to handle
unmodeled factors. Neural Networks can lead to the definition of intelligent
robots as: “being one which trained for a certain class of operations, rather
than one which is trained for virtually all possible applications”.
No comments:
Post a Comment