Improved Gravitational Search Algorithm And Applications

In recent years machine vision / learning methods have achieved better solutions to NP hardproblems. Swarm intelligence optimisation method belongs to intelligence optimisation algorithmwhich basically is about imitating the behaviour of social organisms. Some very notableadvantages of swarm intelligence include: This algorithm does not require a prior knowledge in itsperformance that makes it very easy to implement. The problem is not continuous, the methodneeds no exact mathematical models. The solutions of the problems in SI do not depend onexplicit expression since it can deal with the problem directly, hence, preprocessing is not very necessary.Swarm intelligence is better than traditional optimisation methods because SI can find the solution depending on probability. It further depends on biological behaviour of organisms and modeling the algorithm applying these behavioural patterns is very important in achieving the goals. When SI is implemented appropriately, performance becomes robust, this indicates the better performance of SI algorithms over many traditional methods. Some examples of SI algorithms include: Gravitational Search Algorithm(GSA), Genetic Algorithm(GA), DNA, PSO, Wolf Algorithm(WA), ACO, Bacteria Colony(BC), Fish Algorithm(FA), ABC etc.Furthermore, SI can be divided into single objective and multi-objective population dependent methods. In the aspect of application, SI is applied in: image processing, distributed devices e.g. robotics, to study the algorithm itself etc. With many advantages in SI algorithm, there are also some disadvantages which include: premature convergence, the parameters and particles are usually independent, SI is totally dependent on the environment, SI results are difficult to predict.Feature selection has become a broad area of machine learning, the task of selecting and classifying features is usually daunting, hence, many algorithms to tackle these tasks continue to evolve. One of such developments is in the area of swarm intelligence(SI), SI is a collective performance of a group of social organism e.g. bees, birds, ants, termites, fish, etc., where every member abides by its own existential principles to function in the swarm.This project proposes algorithms mimicking these behavioral patterns of organisms then applied to image feature selection and classification. Hence, to improve the accuracy of image classification; the project is considered to reduce the feature dimension through quantum-binary gravitational search algorithm in the first part. Then support vector machine(SVM) classifier which locates a hyper-plane that gives the minimum number of errors in the process was implemented in ascertaining the efficiency of the method.The second part of the project is Improved Centripetal Accelerated Particle Swarm Optimization(ICAPSO). The original version of CAPSO method has no parameters to suitably regulate the process, so two new parameters were inculcated which improved the efficiency. Not only that; the idea of quantum-behaved particles was applied, but to evaluate the capability of the ICAPSO algorithm, it was tested on medical image database. Considering the aspect of relevance feedback(RF) application in a Content-Based Image Retrieval(CBIR) system, the results showed that ICAPSO was more efficient in the comparison with previous methods.The final part of the work is a multi-objective feature selection with gravitational search algorithm. The goal is to select subsets of available features via elimination of the irrelevant features. To achieve the goals, the high dimensionality was minimised to gain relevant information on the learned models. This method is to maximise the classification performance and minimise the number of features. From the simulation experiment results, the FSMOGSA method is more efficient in comparison with previous methods. The FSMOGSA method reduced the error rate and improved the general performance.