Generative representations for the automated design of modular physical robots

Abstract

1https://ntrs.nasa.gov/search.jsp?R=20030107313 2018-05-13T01:23:13+00:00Zwe have built from this process are still very simple compared to human-engineered machines, their structure is more principled (regular, modular and hierarchical) compared to previously evolved machines of comparable functionality, and the virtual designs which are achieved by the system have an order of magnitude more moving parts. Moreover, we quantitatively demonstrate for this .design space how the generative representation is capable of searching more efficiently than a nongenerative representation. Structure of this paperWe will begin with a brief background of evolutionary robotics and related work, and demonstrate the scaling problem with our own prior results. Next we propose the use of an evolved generative representation as opposed to a non-generative representation. We describe this representation in detail as well as the evolutionary process that uses it. We then compare progress of evolved robots with and without the use of the grammar, and quantify the obtained advantage. Working twodimensional and three-dimensional physical robots produced by the system are shown. BackgroundBiological evolution is characterized as a process applied to a population of individuals, which are subject to selective replication with variation (Maynard-Smith and Szathmary, 1995). Evolutionary design systems use the same principles of biological evolution to achieve machine design, yet add a target to the evolution -the functional requirements specified by the designer. Candidate designs in a population are thus still selected, replicated and varied, but selection is governed by an external design criteria. After a number of generations the selective evolutionary process may breed an acceptable design.Genetic algorithms -a subset of evolutionary computation involving mutation and crossover in a population of fixed length bit strings (Holland, 1975) -have been applied for several decades in many engineering problems as an optimization technique for a fixed set of parameters.Alternatively, more recent open-ended evolutionary design systems, in which the process is allowed to add more and more building blocks and parameters, seem particularly adequate to design problems requiring synthesis. Such open-ended evolutionary design systems have been demonstrated for a variety of simple design problems, including structures, mechanisms, software, optics, robotics, control, and many others (for overviews see. for example, Koza, 1992; Bentley, 4 1999; Husbands et nl, 1998). Yet these accomplishments remain simple compared to what teams of human engineers can design and what nature has produced. The evolutionary design approach is often criticized as scaling badly when challenged with design requirements of higher complexity. (Mataric and Cliff, 1996) While there are still many poorly understood factors that determine the success of evolutionary design -such as starting conditions, variation operators, primitive building blocks and fidelity of simulatio...

Comparing the Evolvability of Generative Encoding Schemes

“…In the large majority of these studies, the performance of evolved individuals is analyzed solely by the walking speed of the robot and the required number of generations of evolution. The rate of evolution and evolved performance has also been linked to evolvability provided by the encoding scheme, wherein controllers achieving a higher task fitness and requiring fewer generations to evolve are considered more evolvable (e.g., see Hornby et al (2003); Komosiński and Rotaru-Varga (2001)). While these approaches provide interesting insights on the characteristic of the underlying genotype-to-phenotype mapping, they largely ignore its capabilities to generate viable phenotypic variations (diverse gaits in case of legged robots).…”

“…Most studies estimate evolvability either as, (i) the proportion of genetic mutations that are beneficial to an individual, irrespective of the phenotypic novelty of the resultant offspring (e.g., Hornby et al (2003); Reisinger and Miikkulainen (2007)), or as (ii) the range and diversity of the phenotypic variants resulting from genetic change (Lehman and Stanley, 2011;Reisinger et al, 2005;Lehman and Stanley, 2013), usually without considering the deleteriousness of the change. Importantly, both these estimates when considered alone do not discount for mutations that, (i) generate very diverse phenotypes but prove lethal to an organism, and (ii) result in minor improvements to a phenotype, but are unable to generate novelty.…”

HybrID: A Hybridization of Indirect and Direct Encodings for Evolutionary Computation

“…However, these direct encodings are limited in their ability to evolve complex, modular, and symmetric phenotypes because individual mutations cannot produce coordinated changes to multiple elements of a phenotype [1]. Such coordinated mutational effects can occur with indirect encodings, also called developmental or generative encodings, wherein a single element in a genotype can influence many parts of the phenotype [1,2].…”

“…However, these direct encodings are limited in their ability to evolve complex, modular, and symmetric phenotypes because individual mutations cannot produce coordinated changes to multiple elements of a phenotype [1]. Such coordinated mutational effects can occur with indirect encodings, also called developmental or generative encodings, wherein a single element in a genotype can influence many parts of the phenotype [1,2]. Indirect encodings have been shown to produce highly regular solutions to problems [1,[3][4][5], but their bias toward regularity makes it difficult for them to properly handle irregularities in problems [4].…”

A novel generative encoding for evolving modular, regular and scalable networks

“…Most such algorithms are generative encodings [17], wherein information in a genotype can influence multiple parts of a phenotype, instead of direct encodings, wherein genotypic information specifies separate aspects of phenotypes. Generative encodings tend to produce more regular phenotypes that outperform direct encodings [3,8,10] and can produce modular and hierarchical structures [10]. Individual generative genomes can also produce networks of various sizes, as appropriate [16,9,7,5].…”