Title

Effect of animat complexity on the evolution of hierarchical control

Abstract

Animal movements are realized by a combination of high-level control from the nervous system and joint-level movement provided by the musculoskeletal system. The digital muscle model (DMM) emulates the low-level musculoskeletal system and can be combined with a high-level artificial neural network (ANN) controller forming a hybrid control strategy. Previous work has shown that, compared to ANN-only controllers, hybrid ANN/DMM controllers exhibit similar performance with fewer synapses, suggesting that some computation is offloaded to the low-level DMM. An open question is how the complexity of the robot, in terms of the number of joints, affects the evolution of the ANN control structure. We explore this question by evolving both hybrid controllers and ANN-only controllers for worm-like animats of varying complexity. Specifically, the number of joints in the worms ranges from 1 to 12. Consistent with an earlier study, the results demonstrate that, in most cases, hybrid ANN/DMM controllers exhibit equal or better performance than ANN-only controllers. In addition, above a threshold for animat complexity (number of joints), the ANNs for one variant of the hybrid controllers have significantly fewer connections than the ANN-only controllers.

Department(s)

Computer Science

Document Type

Conference Proceeding

DOI

https://doi.org/10.1145/3071178.3071246

Keywords

Animats, Artificial neural networks, Controller evolution, Digital muscle model, Evolutionary robotics

Publication Date

7-1-2017

Journal Title

GECCO 2017 - Proceedings of the 2017 Genetic and Evolutionary Computation Conference

Share

COinS