Improving Efficiency of Evolving Robot Designs via Self-Adaptive Learning Cycles and an Asynchronous Architecture
Léni K. Le Goff, Emma Hart · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
Algorithmic frameworks for the joint optimisation of a robot's design and controller often utilise a learning loop nested within an evolutionary algorithm to refine the controller associated with a newly generated robot design. Intuitively, it is reasonable to assume that the length of the learning period required is directly related to the complexity of the new design. Therefore, we propose a novel self-adaptive criterion that modifies the learning budget for each individual robot based on setting a target for the progress to be achieved during learning. This stopping criterion can lead to wide variance in learning times per robot evaluated. Research in other domains where variable evaluation time is also observed has suggested that asynchronous architectures are preferable in this situation, leading to improved objective performance and efficiency. We conduct a systematic comparison of synchronous and asynchronous architectures using the new learning stopping criterion in a joint optimisation task, showing that a judicious choice of target learning progress used in conjunction with an asynchronous framework provides considerably better results in terms of fitness and computational efficiency than a synchronous framework --- in the latter, the choice of target learning progress has no significant influence.