Supplemental Figures for "Enhancing Divergent Search through Extinction Events"

This webpage provides additional figures for the GECCO 2015 paper "Enhancing Divergent Search through Extinction Events," by Joel Lehman and Risto Miikkulainen. Due to limited space, these figures could not be printed in the paper itself, but help to present a complete picture of the results presented therein.

Figure 5 in the paper showed the change of evolvability over evolution for the biped domain for novelty search and objective-based search. In figure S1 below, the corresponding change is shown for the maze domain; in figures S2 and S3, it is shown for behavioral grid search method in the maze and biped domains, respectively.

Figure 6 in the paper showed the dynamics of niche occupation in the biped domain. Figure S4 below shows the same dynamics for the maze domain.

Dynamics of Niche Occupation in the Maze Domain.
Fig.S1 - Change in Evolvability over Evolution in the Maze Domain for Novelty and Objective-based search. The average evolvability of individuals in the population is shown for objective-based search (labeled Obj) and novelty search (labeled Nov) in the maze domain. Extinction setups combined with novelty search result in increased evolvability and an overall increasing trend. Extinction setups combined with objective-based search result in decreased evolvability and an overall stagnating trend. The conclusion is that there is a qualitative difference in how evolvability changes over time between the two search methods.
Fig.S2 - Change in Evolvability over Evolution in the Maze Domain for Behavioral Grid Search. The average evolvability of individuals in the population is shown for behavioral grid search in the maze domain. Evolvability in the Control setup peaks and then declines, while the Extinction setups resist decline. The conclusion is that more frequent extinctions improve evolvability.
Fig.S3 - Change in Evolvability over Evolution in the Biped Domain for Behavioral Grid Search. The average evolvability of individuals in the population is shown for behavioral grid search in the biped domain. As in the maze domain, more frequent extinctions improve evolvability.
Dynamics of Niche Occupation in the Maze Domain.
Fig.S4 - Dynamics of Niche Occupation in the Maze Domain. The average number of niches occupied over evolution is shown for behavioral grid search in the maze domain. The Control setup accumulates niches monotonically, whereas the Extinction setups are decimated at regular intervals. Because each extinction event spares only 10 niches, increasingly quick repopulation in the Extinction setups suggest that the representations are more evolvable in these setups than in the others; a ceiling effect is noticable towards the end of evolution as few niches remain undiscovered. Note that each extinction always brings the model to exactly 10 niches, although this is not visible due to sampling error.