Optimization of Emotional Learning Approach to Control Systems with Unstable Equilibrium - Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing - page 43

of controller itself is a quite time consuming process, even for model free ones like
fuzzy controllers we need to adjust the parameters automatically or manually [4].
Therefore, adaptive and intelligent control are new approaches in control engineer-
ing. In intelligent control there is an immense desire to inspire from natural systems,
by way of illustration, we can refer to Neuro-Fuzzy [7], evolutionary systems [22],
intelligent controllers based on Reinforcement Learning [8], and Multi Agent con-
trollers [23]. Advanced findings in Neuroscience indicate that emotion plays a sig-
nificant role in human reasoning and decision making [18]. Recently, emotion has
also introduced to artificial intelligence as a significant factor in decision making
of expert systems [18]. Humans and other mammals are known as creatures which
have emotional behaviors. The emotional decision making process which usually
takes place fast, helps mammals keep safe in dangerous situations. LIMBIC is a
part of brain in mammals which controls emotional process. This system is been
modeled mathematically in [1][18]. AMYGDALA and ORBITOFRONTAL are two
main parts of limbic, and their models were initially introduced in [19]. Brain Emo-
tional Learning Based Intelligent Controller (BELBIC) introduced as an intelligent
controller [15] in result of attempts and previous works in [19]. BELBIC intelli-
gent controller is developed based on middle brain of mammals called BEL [19].
This controller has been successfully applied to many simulated control problems
[14][17]; moreover, it has been utilized in many real problems, and have had accept-
able results in comparison other control methods[10]. Additionally we can appoint
many other successful applications of this controller such as: HVAC Systems [26],
Robocup issues [24][25], Control of Intelligent Washing Machines [16], Speed Con-
trol of Switched Reluctance Motor [20], Active Queue Management [9], Model Free
Control of Overhead Travelling Crane [10]. The main problem concerning learning
controllers which do not need any background knowledge of system dynamics nor
system models in particular RL based controllers and BELBIC, is that in initial steps
of learning process because of producing false control commands, they might cause
low performance. If this part of learning does not lead to in instability, the controller
can learn the proper control signals progressively otherwise or in the case that the
system is innately unstable, implementing these controllers may cause the system
become unstable and the process must be stopped. Even though BELBIC demon-
strates fast learning ability, it shares the same problem and cannot be applied to
such systems [11]. To overcome similar difficulties another approach is introduced.
In this paper an evolutionary method based on Particle Swarm Optimization (PSO)
[13] is introduced to improve control signal and reduce error gradually. The sensory
signal generation unit is a combination of error and its derivation with constant coef-
ficients. Producing proper stress signal is an important factor in this section. In brief,
in this approach initially, BELBIC controller output must become similar to a pre-
liminary controller. At this level, the stress signal generation unit can be a weighted
sum of some characteristics of error. After early learning, generation of initial signal
will turn to a new weighted sum with 6 weights. The next step is enhancing stress
signal in order to reduce BELBIC controller error in the time of controlling inverse
pendulum. This enhancement will be performed by adjusting the weights of stress
generation neural network through PSO. This approach has demonstrated outstand-