Optimization of Emotional Learning Approach to Control Systems with Unstable Equilibrium - Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing - page 46

There is one more input to Amygdala, A th , which is the maximum values of sensory
inputs:

A th =

( )

max S i (6)

, initial values of A ,
A th , O , and functions S (Sensory Generator) and R (Stress Generator) in producing
emotional signal must be selected properly.

Figure 2 presents BELBIC controller. Ultimately α and β

Fig. 2 Control system configuration using BELBIC

4 Design and Structure of the Controller

Control task of unstable systems or stable systems with unstable equilibrium is one
of most difficult and interesting problems in control issues. Usually, controlling in-
verted pendulum, as one of the problems which is classified in this category, is
utilized for testing and reviewing of proposed approaches in such systems. This
problem is described as tracking reference signal and stabilizing the pendulum at
the same time. Due to nonlinearity of systems state equations, it is not easy to de-
sign a model based controller for this system. In this paper BELBIC is employed
as a model free controller. One of the BELBIC properties is fast learning, however
there is no information about the dynamic of system and the pendulum may falls
down in primary steps; consequently the learning process terminates soon. It means
that BELBIC should learn a proper control signal in the short time regarding to sen-
sory and stress input signals, but due to sensitivity of pendulum angle to error, even
a little error leads system to be unstable. In simulation this leaning process takes too
long and probably impossible in real applications [11].
We have proposed a multi phase solution to solve this problem. In phase one,
an imitative process takes place. BELBIC controller learns the control signal pro-
duced by another basic controller (a proper tuned double PID for this task) which
can stabilize the inverted pendulum. In the next phase, we explore to find the first
solution better than above basic controller ,and finally in the last phase, the solution
of phase two will be improved. The first phase will be finished in short time and net-
works' weights will be updated and fixed. The second phase is designed to find the
first better solution and to replace it with current basic controller, so we the stress
generation method will be changed to the new structure and explores the solution
space. This method employes particle swarm optimization algorithm in order to find
one stress generation block that leads to more error reduction. In this phase stress