TY - CONF T1 - Automated development of latent representations for optimization of sequences using autoencoders T2 - 2021 IEEE Congress on Evolutionary Computation (CEC) Y1 - 2021 A1 - Kaszuba, Piotr A1 - Komosinski, Maciej A1 - Mensfelt, Agnieszka AB - In this paper, we propose an automated method for the development of new representations of sequences. For this purpose, we introduce a two-way mapping from variable length sequence representations to a latent representation modelled as the bottleneck of an LSTM (long short-term memory) autoencoder. Desirable properties of such mappings include smooth fitness landscapes for optimization problems and better evolvability. This work explores the capabilities of such latent encodings in the context of optimization of 3D structures. Various improvements are adopted that include manipulating the autoencoder architecture and its training procedure. The results of evolutionary algorithms that use different variants of automatically developed encodings are compared. JF - 2021 IEEE Congress on Evolutionary Computation (CEC) PB - IEEE UR - http://www.framsticks.com/files/common/LatentRepresentationsForSequencesOptimization.pdf ER - TY - CONF T1 - A Flexible Dissimilarity Measure for Active and Passive 3D Structures and Its Application in the Fitness–Distance Analysis T2 - Applications of Evolutionary Computation Y1 - 2019 A1 - Komosinski, Maciej A1 - Mensfelt, Agnieszka ED - Kaufmann, Paul ED - Castillo, Pedro A. AB - Evolutionary design of 3D structures – either static structures, or equipped with some sort of a control system – is one of the hardest optimization tasks. One of the reasons are rugged fitness landscapes resulting from complex and non-obvious genetic representations of such structures and their genetic operators. This paper investigates global convexity of fitness landscapes in optimization tasks of maximizing velocity and height of both active and passive structures. For this purpose, a new dissimilarity measure for 3D active and passive structures represented as undirected graphs is introduced. The proposed measure is general and flexible – any vertex properties can be easily incorporated as dissimilarity components. The new measure was compared against the previously introduced measure in terms of triangle inequality satisfiability, changes in raw measure values and the computational cost. The comparison revealed improvements for triangle inequality and raw values at the expense of increased computational complexity. The investigation of global convexity of the fitness landscape, involving the fitness–distance correlation analysis, revealed negative correlation between the dissimilarity of the structures and their fitness for most of the investigated cases. JF - Applications of Evolutionary Computation PB - Springer SN - 978-3-030-16692-2 UR - http://www.framsticks.com/files/common/DissimilarityMeasure3DStructuresFitnessDistance.pdf ER -