| dc.description.abstract | In Evolutionary Robotics, Evolutionary Algorithms (EAs) are used to optimize robots. Research has shown that co-optimizing morphology and control can lead to innovative, animal-like behavior. Additionally, co-optimizing in Modular Robotics can be used to automatically produce robots for any task. Even so, it is a definite challenge to design a system that can evolve morphology and control simultaneously. A common issue is that of not being able to properly explore the space of possible robots, and thus not finding the globally best solutions to the task. This is reflected in the field struggling with early convergence of morphology, rugged search landscapes, and overall stagnation. Here, we conduct two different experiments centered around the co-optimization of morphology and control in modular robots. The first investigates different controllers, comparing centralized and decentralized control strategies regarding their effect on morphology and performance. The second experiment investigates gradual encodings and their effect on smoothing the search space. In our gradual encodings, modules grow out gradually instead of being added with full size. We found that a controller that duplicates control units across the robot body performs significantly better than other control approaches because it explores more of the search space. This indicates that decentralization with duplication can be useful, and possibly decrease early convergence of morphology, which helps confirm that compressing the search space is often beneficial. We therefore present this as an argument for duplication in controllers in general. In addition, we found that while the gradual encodings did smooth the search space, they led to no better or worse performance than the baseline. Even though we suggest some instances where it might still be advantageous, it largely implies that there is no benefit to these more gradual encodings in a standard EA. Overall, these two experiments corroborate other research findings that there is a trade-off between fine-tuning and coarsely exploring in a search, and that the latter will often be more helpful initially. We hope that future researchers will benefit from the suggested controller and encoding strategies and our further insights into their effect on search space traversal. | eng |