MEAT: Morphological Evolution of Augmenting Topologies

Abstract

Evolutionary Algorithms (EAs) are effective tools for solving various optimization problems. One of these problems is the co-optimization of robot morphology and control. In modular robotics, this challenge lies in the reconfiguration of the shape of a robot composed of various modular components. With EAs, these configurations can be evolved to create robots able to accomplish a task. Because of the wide range of different configurations, finding the optimal arrangement simultaneously with the control parameters can be very challenging. Even though contemporary approaches are promising, evolutionary runs often lead to premature convergence to suboptimal configurations. To investigate the challenge of premature convergence, this thesis evaluates the performance of evolving modular robots through three different experiments. Experiment 1 investigates how a simple initialization approach compares to a random one, looking at the effects on performance, morphological convergence, morphology, complexity, and diversity. Experiment 2 analyzes the effects of morphological protection, thereby also the performance of combining the simple initialization with protection to create Morphological Evolution of Augmenting Topologies (MEAT). Lastly, Experiment 3 examines how evolving in a challenging environment impacts evolution. The results demonstrate that starting with an initial population without diversity quickly improves and matches methods with higher initial diversity. MEAT performs significantly better in both environments while leading to less complex morphologies. Despite this lower complexity, there are no indications of less exploration of the search space. Maintaining the morphological diversity in the population through morphological protection enables more exploration, and is thereby an effective tool in delaying morphological convergence. These results highlight the potential of MEAT, as augmenting morphological topologies is beneficial when co-optimizing the morphology and control of robots.