Figure 4

Particle swarm optimization vs evolutionary strategy. To compare the two optimization strategies and to show that the non-compactness penalty (d, see Section Cost function) leads in many cases to a better convergence (lower original error term E), we computed Euler diagrams for one previously published gene expression dataset containing genes differentially expressed between a specialized mesenchymal cell type (stellate cells) and normal skin fibroblasts (filter settings: minimum total: 40; maximum total: 140; max p-value: 0.05). For different delta values (= weight of the pressure term) the cost functional E is shown for the two optimization strategies. The number of optimization steps until the stop criterion (maximum number of constant steps = 50, maximum number of steps = 500) is met is shown on the right side. Min/max (light color), interquartile range (dark color), and median value (black line) over n = 20 runs and 20 different delta parameter settings (x-axis) ranging from 0 to 2000 are shown.