Comparison of Different Genotype Encodings for Simulated Three-Dimensional Agents

Abstract

We analyze the effect of different genetic encodings used for evolving three-dimensional agents with physical morphologies. The complex phenotypes used in such systems often require nontrivial encodings. Different encodings used in Framsticks--a system for evolving three-dimensional agents--are presented. These include a low-level direct mapping and two higher-level encodings: one recurrent and one developmental. Quantitative results are presented from three simple optimization tasks (passive height, active height, and locomotion speed). The low-level encoding produced solutions of lower fitness than the two higher-level encodings under similar conditions. Results from recurrent and developmental encodings had similar fitness values but displayed qualitative differences. Desirable advantages and some drawbacks of more complex encodings are established.

Genetic mappings in artificial genomes

“…The range of these representations is wide, and includes approaches dedicated for optimization of control systems (brains) alone (which is traditionally the domain of Arti cial Intelligence) [16,22,14,17,7,10], designs (bodies, morphologies) alone [13,5,9,12,23,25], and both brains and bodies at the same time [26,11,8,1,24,20,21]. However, comparing such representations is di cult because not only genetic encodings are di erent, but also the simulation models vary.…”

“…Some researchers tried to systematize and compare arti cial genetic encodings in a single phenetic model [28,6,20,15]. This allows to test their performance in the same environment, and thus focus on the characteristics of the encodings (features like their complexity, constraints, support for modularity, body symmetry, compression, redundancy, etc.).…”

“…to nd the subset of Y for the given subset of X, and vice versa. repetition gene see [20] for details.…”

“…8 on the next page is analogous to Fig. 6, but it concerns a more complex, developmental genetic encoding, f4 [20]. To achieve such functionality, the two mappings: f4 →f0 and f0 →phenotype must have been computed for the presented f4 genotype, and then composed.…”

“…The freedom in designing and using various genetic encodings allows researchers to compare them and study their properties in a single simulation model [28,6,15,20]. If many genetic encodings exist in a single simulation model, then each of them must be capable of expressing phenotypes in that model.…”