ABSTRACT
This research was aimed at the development of an improved artificial fish swarm optimization algorithm based on knowledge (normative and situational) in cultural algorithm and crossover operator called the modified Cultural Artificial Fish Swarm Algorithm with Crossover (mCAFAC). The Normative and Situational knowledge inherent in cultural algorithm were utilized to guide the step size as well as the direction of evolution of AFSA at different configurations, in order to combat the ease at which AFSA falls into local minima. An inertial weight selection is adopted such that the algorithm can adaptively select its parameters (visual and step size) when searching for global solution. Crossover operator was applied to fuse the AFSA and the modified Cultural Artificial Fish Swarm Algorithm called the mCAFAC, in order to enhance its convergence to a global minimal. Four variations of mCAFAC (mCAFAC_Ns, mCAFAC_Sd, mCAFAC_NsSd and mCAFAC_NsNd) were implemented in Matlab R2013b using different configurations of the cultural knowledge. A total of sixteen test functions (Ackley, Cosine Mixture, Rastrigirn etc.) were employed to evaluate the performance of each mCAFAC variant. Simulation results showed that the mCAFAC outperformed the original AFSA with the mCAFAC_NsSd having superior performance over all the other variants.mCAFAC_NsSd produced the best result in 9 out of the 16 test cases (56.25%) while mCAFAC_NsNd produced the best result in 1 out of the 16 test cases (6.25%), mCAFAC_Ns produced the best result in 3 out of the 16 test cases (18.75%) and mCAFAC_Sd produced the best result in 1 out of the 16 test cases (6.25%). All the variants, including the standard AFSA, the modified AFSA and the replicated ABC produced the same result in 2 out of the 16 test cases (12.5%).mCAFAC_NsSd was then applied to determine the optimal values of the weighting matrices (Q and R) of linear quadruple regulator (LQR) controller.This was validated on the quadrupleInverted Pendulum (QIP) model where the obtained LQR was able to stabilize the model in 7.4161s as against 7.5162s when using the conventional trial-and-error LQR method. This showed a convergence of the solution space using both approaches with the LQR (mCAFAC) having a more optimal time-to-solution.