A Swarm intelligence based approach to optimisation of machine learning problems using a parallelised framework and the cloud

Abstract

In the recent years, the field of machine learning (ML) has become popular and has been utilised to achieve a wide variety of advances in many different areas. However, the information processing capability of certain ML models is dependent on a number of user-defined hyperparameters which need fine tuning for optimal performance. Due to the large sizes of the search spaces and datasets involved, algorithm complexity, long execution times and the high cost of processing, the use of brute force searching is not always a feasible approach. In these cases, finding a near-optimal solution given short time budgets is acceptable.The study describes an approach, based on ant swarm intelligence optimisation, for dealing with these types of combinatorial problems. A parallelisation approach was chosen and applied to the optimisation algorithms with the aim of identifying satisfactory performing ML model parameter configurations in a shorter time period. Finally, an investigation of the benefits and limitations of using cloud virtual machines for the experimentation phases of projects of this type was done during this study. The method has been tested on four classification problems having different sizes and features. The results show that the evolved ML models achieved satisfactory accuracy and generalisation ability while the optimiser applications were able to identify better performing parameter sets during training phase at earlier stages when compared to a purely random search in some cases.