Abstract
This dissertation is about understanding the requirements for successfully implementingTransfer Learning (TL) in the Genetic Algorithms (GA). TL is the procedure of transferring
previous knowledge from an old problem, called the source problem (S) to another problem
called the target problem (T). We have performed this study by implementing the process
of the TL by employing the Genetic Algorithm (GA) as the model solver. GA is a type of
Evolutionary Computation (EC) inspired by biological evolution theory that using biological
evolution strategies by mimicking inheriting characteristics over many generations. TL has some
limitations, for example, negative transfer. This situation halts the performance of solving the
target problem. Also, during our study, we found out transferring the whole final source population to the target problem is not always a beneficial strategy for solving hard or non-related
problems. Our study focuses on understanding the behavior of the transferred population and
how to make them more beneficial to the target solver and the GA.
In this dissertation, we experimented with and evaluated several strategies for transferring
knowledge including the Estimation of Distribution Algorithm (ED). We proposed an algorithm
that samples the transferred population, and we evaluated our algorithm against other strategies
of TL. We experimented and analyzed the effect of the content of the transferred population
on the performance of the target solver. We experimented with transferring partial knowledge
from the source problem to the target problem. We also experimented with sampling and
transferring knowledge from multiple source problems to the target problem. The results of
our studies show how TL can improve the performance of the GA in terms of the number of
generations, time, effort the GA solver took to find the optimal solution. Also, analyzed factors
that affect the GA performance and how to sample transferred population in terms of providing
the GA with needed knowledge from the previous problem.