Project Title: A Metaheuristic Approach to Hyperparameter Optimisation for Classification Models of the Dry Bean Dataset  
Student: Michael CARTER
Course: MSc in Computer Science

Abstract:
The improper tuning of machine learning hyperparameters could result in a machine learning
model failing to achieve its full potential for a given dataset. Hyperparameter optimisation
(HPO) algorithms are often used to reduce the effort and potential for human error that comes
with tuning hyperparameters by hand, however, in the case of the Dry Bean dataset these 
algorithms have been largely overlooked. In this project, the performance of various particle swarm
and genetic HPO algorithms are applied to kNN, SVM, Random Forest and MLP classifiers
with the primary aim of identifying a new best classifier for the dataset. The particle swarm
algorithm with parameter values c1 = 1 and c2 = 2 is identified as the best HPO algorithm in
the study, producing optimal results for 3 out of the 4 classifiers tested, while the MLP classifier
is found to be the most effected by changes in its hyperparameter values. The SVM built in this
study outperformed all other instances of SVMs in the literature with and accuracy of 93.27%,
while the MLP, kNN and random forest models achieved classification accuracies of 93.18%,
92.67%, 92.59% respectively, therefore, a new best classifier of the dataset was not found. It is
suggested that the Bayesian algorithm which was used in the construction of the current best
classifier, may be better suited to machine learning models of this dataset than the particle
swarm and genetic algorithms. It is further suggested that the use of a more complex neural
network coupled with a HPO algorithm may produce a new best classifier in the future.