Metaheuristic optimization algorithms are usually compared by the quality
of the solution they reach, which says nothing about how they searched
for it. This work presents a set of six complementary similarity measures
that compare the behaviour of two algorithms directly from their search
trajectories. Five of the measures are original, while the cosine distance
is adopted from earlier work. We evaluate the measures on 28 algorithms
from the mealpy library, run on the BBOB benchmark suite in dimensions
2, 5 and 10. Spearman correlation shows that the measures are largely only
weakly related, so each contributes a distinct view of behaviour. We further
show that with increasing dimension, search behaviour becomes progressively
decoupled from final solution quality.
|
