Analyzing the Performance of the Linkage Tree Gene-pool Optimal Mixing Evolutionary Algorithm on Tree Decomposition Mk Landscapes using Local Optima Networks
Summary
Abstract
This thesis analyses the performance of Linkage Tree Gene-pool Optimal Mixing Evolutionary
Algorithm (LT-GOMEA) on Tree Decomposition Mk Landscapes for problems
generated by the benchmark function CliqueTreeMk introduced by Thierens et al. [16].
To understand the challenges faced by LT-GOMEA on these landscapes, the search process
is visualized using Local Optima Networks (LONs) introduced by Ochoa et al. [13].
Problems are generated in both a random and deceptive-trap codomain, and the input
parameters for CliqueTreeMk are varied to create problems with different characteristics.
The main parameter being varied is the overlap between subfunctions, and the
number of subfunctions increases as well to maintain a constant problem length. Two
implementations of LT-GOMEA are tested: the standard population-based version and a
non-population-based version in which the building blocks are learned from a hill-climbed
population. For population-based LT-GOMEA, the performance for both codomains increased
prior to decreasing with an increase in overlap, but the deceptive-trap codomain
performed better at higher overlap. This is thought to be due to a lower number of local
optima (than for the random codomain) and decreased deceptiveness with an increase
in overlap. The non-population based LT-GOMEA had similar results for the deceptivetrap
codomain, but performed poorly on the random codomain with overlap higher than
0, suggesting that an evolving population is necessary to utilize the linkage learning of
LT-GOMEA. We also discovered that the intended deceptiveness of the deceptive-trap
codomain decreases with overlap, and this could also contribute to the increase in global
optima for the deceptive-trap neutral codomain. The findings about the challenges faced
by LT-GOMEA are supported by analyzing the search process through the use of LONs.
The LONs reveal that problems generated with the same input parameters can vary significantly
in difficulty, and gaining a deeper understanding of the reasons behind this
variance could improve the effectiveness of CliqueTreeMk as a benchmark function, by
being able to more precisely estimate the difficulty of the generated problems.