Filipe Assunção

Fitness and Novelty in Evolutionary Art

In this paper the effects of introducing novelty search in evolutionary art are explored. Our algorithm combines fitness and novelty metrics to frame image evolution as a multi-objective optimisation problem, promoting the creation of images that are both suitable and diverse. The method is illustrated by using two evolutionary art engines for the evolution of figurative objects and context free design grammars. The results demonstrate the ability of the algorithm to obtain a larger set of fit images compared to traditional fitness-based evolution, regardless of the engine used.

ELICIT: Evolutionary Computation Visualization

ELICIT is a generic tool that enables the visual exploration of evolutionary computation algorithms. It is characterized by the use of simple visual elements to represent information and by the adoption of interactive techniques which allow the navigation between different granularity levels, i.e., it allows the visualization of data from single runs as well as the display of aggregated data, resulting from multiple evolutionary runs. Visualization of lineages is supported, assisting the user in understanding how a given solution was reached. It also provides several visualization modes to inspect the genetic heritage and offspring of individuals and populations. The application was built with the purpose of being capable to deal with different types of representation, allowing the visualization of both genotypes and phenotypes. This paper describes the main visualization modes offered by the tool, presenting examples of its application to tree and graph-based representations.

Graph-Based Evolutionary Art

A graph-based approach for the evolution of Context Free Design Grammars is presented. Each genotype is a directed hierarchical graph and, as such, the evolutionary engine employs graph-based crossover and mutation. We introduce six different fitness functions based on evolutionary art literature and conduct a wide set of experiments. We begin by assessing the adequacy of the system and establishing the experimental parameters. Afterwards, we conduct evolutionary runs using each fitness function individually. Finally, experiments where a combination of these functions is used to assign fitness are performed. Overall, the experimental results show the ability of the system to optimize the considered functions, individually and combined, and to evolve images that have the desired visual characteristics.

Towards the evolution of multi-layered neural networks: a dynamic structured grammatical evolution approach

Current grammar-based NeuroEvolution approaches have several shortcomings. On the one hand, they do not allow the generation of Artificial Neural Networks (ANNs) composed of more than one hidden-layer. On the other, there is no way to evolve networks with more than one output neuron. To properly evolve ANNs with more than one hidden-layer and multiple output nodes there is the need to know the number of neurons available in previous layers. In this paper we introduce Dynamic Structured Grammatical Evolution (DSGE): a new genotypic representation that overcomes the aforementioned limitations. By enabling the creation of dynamic rules that specify the connection possibilities of each neuron, the methodology enables the evolution of multi-layered ANNs with more than one output neuron. Results in different classification problems show that DSGE evolves effective single and multi-layered ANNs, with a varying number of output neurons.

Automatic generation of neural networks with structured Grammatical Evolution

The effectiveness of Artificial Neural Networks (ANNs) depends on a non-trivial manual crafting of their topology and parameters. Typically, practitioners resort to a time consuming methodology of trial-and-error to find and/or adjust the models to solve specific tasks. To minimise this burden one might resort to algorithms for the automatic selection of the most appropriate properties of a given ANN. A remarkable example of such methodologies is Grammar-based Genetic Programming. This work analyses and compares the use of two grammar-based methods, Grammatical Evolution (GE) and Structured Grammatical Evolution (SGE), to automatically design and configure ANNs. The evolved networks are used to tackle several classification datasets. Experimental results show that SGE is able to automatically build better models than GE, and that are competitive with the state of the art, outperforming hand-designed ANNs in all the used benchmarks.

Evolving the Topology of Large Scale Deep Neural Networks

In the recent years Deep Learning has attracted a lot of attention due to its success in difficult tasks such as image recognition and computer vision. Most of the success in these tasks is merit of Convolutional Neural Networks (CNNs), which allow the automatic construction of features. However, designing such networks is not an easy task, which requires expertise and insight. In this paper we introduce DENSER, a novel representation for the evolution of deep neural networks. In concrete we adapt ideas from Genetic Algorithms (GAs) and Grammatical Evolution (GE) to enable the evolution of sequences of layers and their parameters. We test our approach in the well-known image classification CIFAR-10 dataset. The results show that our method: (i) outperforms previous evolutionary approaches to the generations of CNNs; (ii) is able to create CNNs that have state-of-the-art performance while using less prior knowledge (iii) evolves CNNs with novel topologies, unlikely to be designed by hand. For instance, the best performing CNN obtained during evolution has an unexpected structure using six consecutive dense layers. On the CIFAR-10 the best model reports an average error of 5.87% on test data.

Structured Grammatical Evolution: A Dynamic Approach

Grammars have attracted the attention of researchers within the Evolutionary Computation field, specially from the Genetic Programming community. The most successful example of the use of grammars by GP is Grammatical Evolution (GE). In spite of being widely used by practitioners of different fields, GE is not free from drawbacks. The ones that are most commonly pointed out are those linked with redundancy and locality of the representation. To address these limitations Structured Grammatical Evolution (SGE) was proposed, which introduces a one-to-one mapping between the genotype and the non-terminals. In SGE the input grammar must be pre-processed so that recursion is removed, and the maximum number of expansion possibilities for each symbol determined. This has been pointed out as a drawback of SGE and to tackle it we introduce Dynamic Structured Grammatical Evolution (DSGE). In DSGE there is no need to pre-process the grammar, as it is expanded on the fly during the evolutionary process, and thus we only need to define the maximum tree depth. Additionally, it only encodes the integers that are used in the genotype to phenotype mapping, and grows as needed during evolution. Experiments comparing DSGE with SGE show that DSGE performance is never worse than SGE, being statistically superior in a considerable number of the tested problems.

DENSER: deep evolutionary network structured representation

Deep evolutionary network structured representation (DENSER) is a novel evolutionary approach for the automatic generation of deep neural networks (DNNs) which combines the principles of genetic algorithms (GAs) with those of dynamic structured grammatical evolution (DSGE). The GA-level encodes the macro structure of evolution, i.e., the layers, learning, and/or data augmentation methods (among others); the DSGE-level specifies the parameters of each GA evolutionary unit and the valid range of the parameters. The use of a grammar makes DENSER a general purpose framework for generating DNNs: one just needs to adapt the grammar to be able to deal with different network and layer types, problems, or even to change the range of the parameters. DENSER is tested on the automatic generation of convolutional neural networks (CNNs) for the CIFAR-10 dataset, with the best performing networks reaching accuracies of up to 95.22%. Furthermore, we take the fittest networks evolved on the CIFAR-10, and apply them to classify MNIST, Fashion-MNIST, SVHN, Rectangles, and CIFAR-100. The results show that the DNNs discovered by DENSER during evolution generalise, are robust, and scale. The most impressive result is the 78.75% classification accuracy on the CIFAR-100 dataset, which, to the best of our knowledge, sets a new state-of-the-art on methods that seek to automatically design CNNs.

EvoFashion: Customising Fashion Through Evolution

In today’s society, where everyone desires unique and fashionable products, the ability to customise products is almost mandatory in every online store. Despite of many stores allowing the users to personalize their products, they do not always do it in the most efficient and user-friendly manner. In order to have products that reflect the user’s design preferences, they have to go through a laborious process of picking the components that they want to customise. In this paper we propose a framework that aims to relieve the design burden from the user side, by automating the design process through the use of Interactive Evolutionary Computation (IEC). The framework is based on a web-interface that facilitates the interaction between the user and the evolutionary process. The user can select between two types of evolution: (i) automatic; and (ii) partially-automatic. The results show the ability of the framework to promote evolution towards solutions that reflect the user aesthetic preferences.

Using GP Is NEAT: Evolving Compositional Pattern Production Functions

The success of Artificial Neural Networks (ANNs) highly depends on their architecture and on how they are trained. However, making decisions regarding such domain specific issues is not an easy task, and is usually performed by hand, through an exhaustive trial-and-error process. Over the years, researches have developed and proposed methods to automatically train ANNs. One example is the HyperNEAT algorithm, which relies on NeuroEvolution of Augmenting Topologies (NEAT) to create Compositional Pattern Production Networks (CPPNs). CPPNs are networks that encode the mapping between neuron positions and the synaptic weight of the ANN connection between those neurons. Although this approach has obtained some success, it requires meticulous parameterisation to work properly. In this article we present a comparison of different Evolutionary Computation methods to evolve Compositional Pattern Production Functions: structures that have the same goal as CPPNs, but that are encoded as functions instead of networks. In addition to NEAT three methods are used to evolve such functions: Genetic Programming (GP), Grammatical Evolution, and Dynamic Structured Grammatical Evolution. The results show that GP is able to obtain competitive performance, often surpassing the other methods, without requiring the fine tuning of the parameters.