Morpho-evolution with learning using a controller archive as an inheritance mechanism

TY - JOUR

T1 - Morpho-evolution with learning using a controller archive as an inheritance mechanism

AU - Le Goff, Léni

AU - Buchanan Berumen, Edgar

AU - Hart, Emma

AU - Eiben, A. E.

AU - Li, Wei

AU - De Carlo, Matteo

AU - Winfield, Alan

AU - Hale, Matthew

AU - Woolley, Robert

AU - Timmis, Jon

AU - Tyrrell, Andy

N1 - © 2021 IEEE. This is an author-produced version of the published paper. Uploaded in accordance with the publisher’s self-archiving policy. Further copying may not be permitted; contact the publisher for details

PY - 2023/6/1

Y1 - 2023/6/1

N2 - The joint optimisation of body-plan and control viaevolutionary processes can be challenging in rich morphologicalspaces in which offspring can have body-plans that are verydifferent from either of their parents. This causes a potential mismatchbetween the structure of an inherited controller and thenew body. To address this, we propose a framework that combinesan evolutionary algorithm to generate body-plans and a learningalgorithm to optimise the parameters of a neural controller. Thetopology of this controller is created once the body-plan of eachoffspring body-plan is generated. The key novelty of the approachis to add an external archive for storing learned controllers thatmap to explicit ‘types’ of robots (where this is defined with respectthe features of the body-plan). By learning from a controller withan appropriate structure inherited from the archive, rather thanfrom a randomly initialised one, we show that both the speedand magnitude of learning increases over time when comparedto an approach that starts from scratch, using three differenttest-beds. The framework also provides new insights into thecomplex interactions between evolution and learning, and therole of morphological intelligence in robot design.

AB - The joint optimisation of body-plan and control viaevolutionary processes can be challenging in rich morphologicalspaces in which offspring can have body-plans that are verydifferent from either of their parents. This causes a potential mismatchbetween the structure of an inherited controller and thenew body. To address this, we propose a framework that combinesan evolutionary algorithm to generate body-plans and a learningalgorithm to optimise the parameters of a neural controller. Thetopology of this controller is created once the body-plan of eachoffspring body-plan is generated. The key novelty of the approachis to add an external archive for storing learned controllers thatmap to explicit ‘types’ of robots (where this is defined with respectthe features of the body-plan). By learning from a controller withan appropriate structure inherited from the archive, rather thanfrom a randomly initialised one, we show that both the speedand magnitude of learning increases over time when comparedto an approach that starts from scratch, using three differenttest-beds. The framework also provides new insights into thecomplex interactions between evolution and learning, and therole of morphological intelligence in robot design.

KW - Evolutionary robotics

KW - Embodied Intelligence

U2 - 10.1109/TCDS.2022.3148543

DO - 10.1109/TCDS.2022.3148543

M3 - Article

SN - 2379-8920

VL - 15

SP - 507

EP - 517

JO - IEEE Transactions on Cognitive and Developmental Systems

JF - IEEE Transactions on Cognitive and Developmental Systems

IS - 2

M1 - 15

ER -