Compositional pattern-producing network

Abstract: In this work we introduce a differentiable version of the Compositional Pattern Producing Network, called the DPPN. Unlike a standard CPPN, the topology of a DPPN is evolved but the weights are learned. A Lamarckian algorithm, that combines evolution and learning, produces DPPNs to reconstruct an image. Our main result is that DPPNs can be evolved/trained to compress the weights of a denoising autoencoder from 157684 to roughly 200 parameters, while achieving a reconstruction accuracy comparable to a fully connected network with more than two orders of magnitude more parameters. The regularization ability of the DPPN allows it to rediscover (approximate) convolutional network architectures embedded within a fully connected architecture. Such convolutional architectures are the current state of the art for many computer vision applications, so it is satisfying that DPPNs are capable of discovering this structure rather than having to build it in by design. DPPNs exhibit better generalization when tested on the Omniglot dataset after being trained on MNIST, than directly encoded fully connected autoencoders. DPPNs are therefore a new framework for integrating learning and evolution.

Abstract: This demonstration presents an implementation of a computer-assisted approach to music generation called functional sca↵olding for musical composition (FSMC) whose representation facilitates creative combination, exploration, and transformation of musical ideas and spaces. The approach is demonstrated through a program called MaestroGenesis with a convenient GUI that makes it accessible to even non- musicians. Music in FSMC is represented as a functional relationship between an existing human composition, or sca↵old ,a nd a generated accompaniment. This relationship is represented by a type of artificial neural network called a compositional pattern producing network (CPPN). A human user without any musical expertise can then explore how accompaniment can relate to the sca↵old through an interactive evolutionary process akin to animal breeding.

Abstract: The evolutionary method is an approach to the difficulties of designing soft-bodied robots. One of the prominent methods is compositional pattern producing network with neuroevolution of augmenting topologies (CPPN-NEAT). How-ever, previous research has focused on single-function robots, and the design of multi-functional robots is still unsolved. This study provides a method for generating multi-functional robots by combining the genotype networks of single-functional robots in a modular manner. The proposed method includes the addition of a weight layer during network combination and the selection of populations with a fitness estimator. We conducted experiments to design voxel-based creatures that can perform two types of tasks in the simulation. Target tasks include terrestrial and aquatic locomotion. The results show that the proposed method was able to search for a form that satisfied the two tasks simultaneously faster than the existing methods. Observations of the generated populations indicated that the proposed method enables the efficient exploration of body morphology. Further, a modularized combination helps focus the exploration in a feasible morphology space. Finally, we fabricated evolved soft creatures in the real world as soft-bodied robots by limiting the arrangement of actuation voxels. We believe that the proposed method of designing a multi-functional robot while utilizing existing single-functional robots will contribute to the automatic design of multi-functional soft robots.

Abstract: vii