ABSTRACT. Exploring the potentialities of artificial intelligence (AI) in the world of arts is fundamental to understand and define how this technology is shaping our creativity. In this work, we propose a system that generates emotionally expressive paintings from EEG signals. The emotional information is encoded from the signals through a graph neural network and it is inputted to a generative adversarial network (GAN), trained on a dataset of paintings. The design and experimental choices at the base of this work rely on the understanding that emotions are hard to define and formalize. Despite this, the proposed results demonstrate how the interaction between an AI system and a human can produce an original and artistic re-interpretation of emotions, that have a promising potential for the application of AI technologies to visual arts.

ABSTRACT. Typography is considered by many authors to be the visual manifestation of language. Over time, many designers explore connections between type design and sound, trying to bridge the gap between the two areas. This paper describes \textit{SpeechTyper}, an ongoing system that generates typographic compositions based on speech. Our goal is to create typographic representations that convey aspects of oral communication expressively. The system takes a pre-processed analysis of speech recordings and uses it to affect the glyph design of the recited words. The glyphs' structure is generated using a system we developed previously that extracts skeletons from existing typefaces.

ABSTRACT. Graphic Design (GD) artefacts aim to attract people’s attention before any forward objectives. Thus, one of the goals of GD is frequently finding disruptive aesthetics that stand out over competing design artefacts (such as other books covers in a store or other posters on the street). However, as GD is increasingly being democratized and broadly shared through social media, designers tend to adopt trendy solutions, lacking disruptive and catchy visual features. EvoDesigner aims to assist the exploration of innovative graphic design solutions by using an automatic evolutionary approach to evolve the design of a number of text, shapes, and image elements inside two-dimensional canvases (pages). To enable the collaboration human-machine, the process has been integrated into Adobe inDesign, so human designers and EvoDesigner may alternately edit and evolve the same design projects, using the same desktop-publishing software. In this paper, an overview of the proposed system is presented along with the experimental setup and results accomplished so far on an evolutionary engine developed. The results suggest the viability of the development made in this first iteration of the system, which aims to reinterpret existing layouts in an unexpected manner.

ABSTRACT. A fixed number of brush strokes images are initialized on a canvas, their position, size, rotation, colour, stroke type and drawing index all randomly chosen. These attributes are then modifed by stochastic hillClimbing, simulated annealing or the plant propagation algorithm, approximating a target image ever closer. Simulated annealing showed the best performance, followed by hillClimbing; the plant propagation algorithm performed worst. Finally, the distribution of the attributes of the brush strokes shows us that there appears to be a preference for smaller brush strokes, and strokes of the fourth type.

ABSTRACT. Evolutionary algorithms (ES) have been used in the digital art scene since the 1970s. A popular application of genetic algorithms is to optimize the procedural placement of vector graphic primitives to resemble a given painting. In recent years, deep learning-based approaches have also been proposed to generate procedural drawings, which can be optimized using gradient descent. In this work, we revisit the use of evolutionary algorithms for computational creativity. We find that modern ES algorithms, when tasked with the placement of shapes, offer large improvements in both quality and efficiency compared to traditional genetic algorithms, and even comparable to gradient-based methods. We demonstrate that ES is also well suited at optimizing the placement of shapes to fit the CLIP model, and can produce diverse, distinct geometric abstractions that are aligned with human interpretation of language.

ABSTRACT. We present experiments in solving constrained combinatorial optimization problems by means of Quantum Annealing. We describe how to model a hard combinatorial problem, the Costas Array Problem, in terms of QUBO (Quadratic Unconstrained Binary Optimization). QUBO is the input language of quantum computers based on quantum annealing such as the D-Wave systems and of the ”quantum-inspired” special-purpose hardware such as Fujitsu’s Digital Annealing Unit or Hitachi’s CMOS Annealing Machine. We implemented the QUBO model for the Costas Array Problem on several hardware solvers based on quantum annealing (D-Wave Advantage, Fujitsu DA3 and Fixstars AE) and present preliminary performance result for these implementations, which are also compared with state-of-the-art metaheuristics solvers on classical hardware.

ABSTRACT. Quantum and Quantum-inspired optimisation is becoming increasingly popular. Some of the well-known solvers in this category use local search techniques that share some similarities with Simulated Annealing. Examples are the Digital Annealer and Quantum Annealer. Quadratic Unconstrained Binary Optimisation (QUBO) or Ising Model is a common formulation used by these solvers.

There are many combinatorial optimisation problems that are naturally represented as permutations. An example is the well-known Travelling Salesman Problem (TSP). Converting permutation problems into QUBO entails using an alternative representation which is of binary form. This process entails penalising solutions that cannot be decoded as a valid permutation. Encoding the problem as `on' and `off' bits only (i.e. 0 and 1 in QUBO or -1 and 1 in Ising Model) significantly increases the problem size and the search space. Commercial QUBO solvers are however designed to use specialised hardware to achieve enormous speed up.

An important task when converting permutation problems to QUBOs is that of setting penalty weights. The process of setting static penalty weights (within the context of QUBO formulations) without additional domain knowledge is not trivial. This is because values that are too small will lead to infeasible solutions being returned by the solver while values that are too large may lead to slower convergence. In this study, we explore some methods of setting penalty weights within the context of QUBO formulations. We propose new static methods of calculating penalty weights which lead to more promising results than existing methods.

ABSTRACT. Two mechanisms have recently been proposed that can significantly speed up finding distant improving solutions via mutation, namely using a random mutation rate drawn from a heavy-tailed distribution (``fast mutation'', Doerr et al. (2017)) and increasing the mutation strength based on stagnation detection (Rajabi and Witt (2020)). Whereas the latter can obtain the asymptotically best probability of finding a single desired solution in a given distance, the former is more robust and also performs well when many improving solutions in a larger distance exist.

In this work, we propose a mutation strategy that combines ideas of both two mechanisms. We show that it also obtains the best possible probability of finding a single distant solution. However, when several improving solutions exist, it can outperform both the stagnation-RLS approach and fast mutation. The new operator outperforms is more than an interleaving of the two previous mechanisms and also outperforms any such interleaving.

Our experiments confirm that the asymptotic gains are visible also for realistic problem sizes.

ABSTRACT. The shortest common supersequence problem (SCSP) is a well-known NP-hard problem with many applications, in particular in data compression, computational molecular biology, and text editing. It aims at finding for a given set of input strings a shortest string such that every string from the set is a subsequence of the computed string. Due to its NP-hardness, many approaches have been proposed to tackle the SCSP heuristically. The currently best-performing one is based on beam search (BS). In this paper, we present a novel heuristic (AEL) for guiding a BS, which approximates the expected length of an SCSP of random strings, and embed the proposed heuristic into a multilevel probabilistic beam search (MPBS). To overcome the arising scalability issue of the guidance heuristic, a cut-off approach is presented that reduces large instances to smaller ones. The proposed approaches are tested on two established sets of benchmark instances. MPBS guided by AEL outperforms the so far leading method on average on a set of real instances. For many instances new best solutions could be obtained.

ABSTRACT. We describe a new method for the vehicle routing problems with constraints. Instead of trying to improve the typical metaheuristics used to efficiently solve vehicle routing problems, like large neighborhood search, iterated local search, or evolutionary algorithms, we allow them to explore the deeply infeasible regions of the search space in a controlled way. The key idea is to find solutions better in terms of the objective function even at the cost of violation of constraints, and then try to restore feasibility of the obtained solutions at a minimum cost. Furthermore, in order to preserve the best feasible solutions, we maintain two diversified pools of solutions, the main pool and the temporary working pool. The main pool stores the best diversified (almost) feasible solutions, while the working pool is used to generate new solutions and is periodically refilled with disturbed solutions from the main pool. We demonstrate our method on the vehicle routing problems, with variants respecting time, vehicle capacity and fleet limitation constraints. Our method provided a large number of new best-known solutions on well-known benchmark datasets. Although our method is designed for the family of vehicle routing problems, its concept is fairly general and it could potentially be applied to other NP-hard problems with constraints.

ABSTRACT. Vibrational spectroscopy can be used for rapid nutrient assessment of horticultural produce as a means of quality control. Most commonly, spectral data are calibrated against chemical reference data, which are acquired through resource-intensive analytical methods, using partial least squares regression (PLSR). Recently, genetic algorithms (GAs) have been applied to assist PLSR to construct high-performing models through feature selection and latent variable selection. The current approach relies on manually pre-processed data, which requires human expertise and produces inherent biases. To address this limitation, this paper aims to develop a new GA method for automatically selecting the most appropriate pre-processing techniques for specific tasks to bypass manual pre-processing. The results for infrared spectroscopy show the potential of this approach in out-performing manual pre-processing, while the raman spectroscopy results are competitive, which demonstrates the utility of the approach in terms of saving time and resources.

ABSTRACT. The continuous evolution of economy hinders the decision-making process in this field. The former requires sophisticated techniques, and thus, manual and empirical methods are becoming increasingly obsolete. In this paper, we propose a computationally-supported approach that performs an economic profiling of cities based on their economic features and also a prediction of the future evolution of these economical metrics. Our contributions include (I) a data-ingestion module to extract, transform and load data, (II) a profiling module that achieves an unsupervised classification via a new distance-based cellular genetic algorithm and K-means and (III) a prediction unit based on long short-term memory artificial neural networks. Our proposal is tested on Spain, analysing all its 52 cities, where we use 33 types of real-world economic data that have been recorded monthly for fifteen years. All data has been obtained from the Spanish National Institute of Statistics. Our experiments show that the 52 cities could be clustered into only 3 economical profiles. This decrease in the complexity of entities to be considered allows managers at several levels and countries to take faster and accurate decisions by dealing with few profiles rather than treating each city apart. Also, we found that each profile contains repetitive similarity patterns that are not only determined by economics but also indirectly ruled by the cities' geo and demographic situations. Results also showed our prototype's promising economic predictions

ABSTRACT. Federated learning is a training paradigm according to which a server-based model is cooperatively trained using local models running on edge devices and ensuring data privacy. These devices exchange information that induces a substantial communication' load, which jeopardises the functioning efficiency. The difficulty of reducing this overhead stands in achieving this without decreasing the model's efficiency (contradictory relation). To do so, many works investigated the compression of the pre/mid/post-trained models and the communication rounds, separately, although they jointly contribute to the communication overload. Our work aims at optimising communication overhead in federated learning by (I) modelling it as a multi-objective problem and (II) apply the non-dominated sorting genetic algorithm II for solving it. As far as the authors' know, this is the first work that (I) explores the add-in that evolutionary computation could bring for solving such a problem and (II) considers both the neurone's and devices features together. We perform the experimentation by simulating server/client architecture with 4 slaves. We investigate both convolutional and fully-connected neural networks with 12 and 3 layers, 887,530 and 101,770 weights, respectively. We conducted the validation on the MNIST dataset containing 70,000 images. The experiments have showed that our proposal could reduce the communication by 99% and maintain an accuracy equal to the one obtained by a FedAvg Algorithm that uses 100% of communications.

ABSTRACT. In a free market, the creation of hospitals, schools, sports and public residential facilities, requires the expertise -and possibly the capital- of the private sector. The traditional contract, in which the public administration pays private operators to make or maintain buildings and services, is flanked by public private partnership, in which the private operator is usually delegated to carry out the entire process receiving a fixed fee. For years, governments and administrations have been incentivized to use this kind of contract, assuming that it would increase the building qualities and reduce the risk of higher expenses. Empirical evidence refutes this assumption, and this can be caused by to the so-called moral hazard of the private operator. One of the main problem in public private partnership is the difficulty to define an optimal risk allocation, as there no formulas exist to simulate the performance of the contract in advance. In this paper, Evolutionary Algorithms are used to compute an optimal specifications document, while, at the same time, foreseeing an optimal effort in work. Experimental results clearly demonstrate the feasibility of this approach, also helping the public administration to check if their knowledge is sufficient to structure an efficient specifications document.

ABSTRACT. We propose SonOpt, the first (open source) data sonification application for monitoring the progress of bi-objective population-based optimization algorithms during search, to facilitate algorithm understanding. SonOpt provides insights into convergence/stagnation of search, the evolution of the approximation set shape, location of recur-ring points in the approximation set, and population diversity. The benefits of data sonification have been shown for various non-optimization related monitoring tasks. However, very few attempts have been made in the context of optimization and their focus has been exclusively on single-objective problems. In comparison, SonOpt is designed for bi-objective optimization problems, relies on objective function values of non-dominated solutions only, and is designed with the user (listener)in mind to reduce fatigue, avoid convolution of multiple sounds, ands peed of getting used to the system. This is achieved using two sonification paths relying on the concepts of wavetable and additive synthesis. This paper motivates and describes the architecture of SonOpt, and then validates SonOpt for two popular multi-objective optimization algorithms (NSGA-II and MOEA/D).

ABSTRACT. This paper presents a novel approach to guide a Generative Adversarial Network trained on the FashionGen dataset to generate designs corresponding to target fashion styles. Finding the latent vectors in the generator's latent space that correspond to a style is approached as an evolutionary search problem. A Gaussian mixture model is applied to identify fashion styles based on the higher-layer representations of outfits in a clothing-specific attribute prediction model. Over generations, a genetic algorithm optimizes a population of designs with the objective of increasing their probability of belonging to one of the Gaussian mixture components, or styles. The developed system is able to generate images of maximum fitness visually resembling certain styles. However, some generated designs reveal that the fitness measure relies on a machine-specific understanding of style. Further investigation into the interplay between style model and the exploration of the latent space as well as into the parameter setting of the genetic algorithm is required to align the results with a human understanding of style.

ABSTRACT. Interactive storytelling, is a form of digital entertainment that is gaining great attention over the time. However, one of the main problems this field has, is the poor control that the content creator has over the user once the story starts. Hence, we propose to use artificial intelligence to delegate the creative control to the content creator by designing a system that guides the user's emotions over the story. Specifically, we have developed an EEG-based emotion recognition system trained on EEG recordings acquired from five participants while watching 384 carefully selected videos. This system can do a binary classification on both valence and arousal with an accuracy of 62% and 57%, respectively. We have also made a short-film where each scene adapts to the user's emotion based on predefined interactions established by the content creator. Results have shown that this system does not only improve the engagement of the user, but it also induces an emotion that is closer to the one that the content creator defined for the story. The results of this study indicate that there is a practical application of emotional-based studies for future content creators to better control an intended emotional response delivered and received from the audience.

ABSTRACT. In this paper, we introduce an inverse kinematics model for a robotic manipulator using Genetic Programming (GP). The underlying problem requires learning of multiple joint parameters of the manipulator to reach a desired position in the Cartesian space. We present a new approach to identify a closed-form solution for the Inverse Kinematics (IK) problem, namely IK-CCGP. The novelty of IK-CCGP is the cooperative coevolutionary learning strategy. Unlike other GP approaches, IK-CCGP is not limited to a certain angle combination to reach a given pose and is designed to achieve more flexibility in the learning process. Moreover, it can operate both as single- and multi-objective variants. In this paper, we investigate whether the inclusion of further objectives, i.e. correlation and the consistency of a solution with physical laws, contributes to the search process. Our experiments show that the combination of the two objectives, error and correlation, performs very well for the given problem and IK-CCGP performs the best on a kinematic unit of two joints. While our approach cannot attain the same accuracy as Artificial Neural Networks, it overcomes the explainability gap of IK models developed using ANNs.

ABSTRACT. Cyber-Physical Systems (CPS) are prevalent in critical infrastructures and a prime target for cyber-attacks. Multivariate time series data generated by sensors and actuators of a CPS can be monitored for detecting cyber-attacks that introduce anomalies in those data. We use Signal Temporal Logic (STL) formulas to tightly describe the normal behavior of a CPS, identifying data instances that do not satisfy the formulas as anomalies. We learn an ensemble of STL formulas based on observed data, without any specific knowledge of the CPS being monitored. We propose an algorithm based on Grammar-Guided Genetic Programming (G3P) that learns the ensemble automatically in a single evolutionary run. We test the effectiveness of our data-driven proposal on two real-world datasets, finding that the proposed one-shot algorithm provides good detection performance.

ABSTRACT. Geometric semantic genetic programming (GSGP) is a well-known variant of genetic programming (GP) where recombination and mutation operators have a clear semantic effect. Both kind of operators have randomly selected parameters that are not optimized by the search process. In this paper we combine GSGP with a well-known gradient-based optimizer, Adam, in order to leverage the ability of GP to operate structural changes of the individuals with the ability of gradient-based methods to optimize the parameters of a given structure.

Two methods, named HYB-GSGP and HeH-GSGP, are defined and compared with GSGP on a large set of regression problems, showing that the use of Adam can improve the performance on the test set. The idea of merging evolutionary computation and gradient-based optimization is a promising way of combining two methods with very different -- and complementary -- strengths.

ABSTRACT. Genetic programming is a method to generate computer programs automatically for a given set of input/output examples that define the user's intent. In real-world software development this method could also be used, as a programmer could first define the input/output examples for a certain problem and then let genetic programming generate the functional source code. However, a prerequisite for using genetic programming as support system in real-world software development is a high performance and generalizability of the generated programs. For some program synthesis benchmark problems, however, the generalizability to previously unseen test cases is low especially when lexicase is used as parent selection method. Therefore, we combine in this paper lexicase selection with small batches of training cases and study the influence of different batch sizes on the program synthesis performance and the generalizability of programs generated with genetic programming. For evaluation, we use three common program synthesis benchmark problems. We find that the selection pressure can be reduced even when small batch sizes are used. Moreover, we find that, compared to standard lexicase selection, the obtained success rates on the test set are similar or even better when combining lexicase with small batches. Furthermore, also the generalizability of the found solutions can often be improved.

ABSTRACT. We present a method for discriminating between healthy subjects and Alzheimer’s diseases patients from on-line handwriting. Departing from the current state of the art methods, that adopts machine learning methods and tools for building the classifier, we propose to apply the Negative Selection Algorithm. The major advantage of the proposed method in comparison with others machine learning techniques is that it requires only data by healthy subjects to build the classifier, thus avoiding to collect patient data, as requested by competing techniques. Experiments results involving data produced by 175 subjects show that the proposed method achieves state-of-the-art performance.

ABSTRACT. This paper considers a swarm optimisation approach to few-view tomographic reconstruction. DFOMAX, a high diversity swarm op-timiser, demonstrably reconstructs binary images to a high fidelity, out-performing a leading algebraic technique, differential evolution and particle swarm optimisation on four standard phantoms. The paper considers the effectiveness of optimisers that have been developed for optimal low dimensional performance and concludes that trial solution clamping on the walls of the feasible search space is important for good performance.

ABSTRACT. Cognitive impairments affect millions of persons worldwide, and especially elderly ones. These impairments may be one of the first signs of the arising of neurodegenerative diseases, such as Alzheimer's and Parkinson's, and it is expected that the incidence of this kind of diseases will dramatically increase worldwide in the near future. For this reason, the improvement of the tools currently used to diagnose these diseases is becoming crucial. Handwriting is one of the human skills affected by this kind of impairments, and anomalies such as micrographia have been adopted as diagnosis sign for the Parkinson's disease. In a previous paper, we presented a study in which the handwriting of the subjects involved was recorded while they were performing some elementary tasks, such as the copy of simple words or the drawing of elementary forms. Then we extracted the features characterizing the dynamics of the handwriting and used them to train a classifier to predict whether the subject analyzed was affected by a cognitive impairment or not. In this paper, we present a system that uses vectorial Genetic Porgramming for . The experimental results confirmed the effectiveness of the proposed approach

ABSTRACT. Spinal cord atrophy is one of the neuroimaging features associated with neurodegenerative diseases, inflammatory diseases and trauma. Segmentation methods with varying degree of manual intervention, reproducibility and accuracy have been proposed to assess cord atrophy from MR images, especially at C2 vertebral level. Manual outlining from experts, digital phantoms or MR images of phantoms with known dimension are the only existing ground-truth segmentations for methods comparison. Yet, phantoms images are insufficiently realistic to represent the complexity of the spinal cord MR images, especially in the pathological context. Manual outlining is time consuming, subject to intra- and inter-observer variability, and its accuracy highly depends on the operator’s experience. In this context, there is a clear need for methods that simplifies and facilitates expert intervention, while providing an accurate quantification of cord atrophy. In this paper, we build and test a ground-truth segmentation approach based on a simple evolutionary algorithm (EAcord). EAcord integrates a set of segmentation methods with varying accuracy and manual intervention (manual, semi-automated and automated methods), as well as knowledge about the spinal cord anatomy, its relative location, immediate surrounding environment and shape at C2 vertebral level. A lighter version of EAcord, EAcord-light, is also proposed, using only segmentations from semi-automated and automated methods as inputs. An experimental analysis of EAcord and EAcord-light showed an improved reproducibility and similar to better ground-truth accuracy value than manual outlining. More interestingly, in some cases, EAcord-light is able to produce a ground-truth segmentation with minimal expert intervention.

ABSTRACT. We introduce a new system for data-driven audio sound model design built around two different neural network architectures, a Generative Adversarial Network (GAN) and a Recurrent Neural Network (RNN), that takes advantage of the unique characteristics of each to achieve the system objectives that neither is capable of addressing alone. The objective of the system is to generate interactively controllable sound models given (a) a range of sounds the model should be able to synthesize, and (b) a specification of the parametric controls for navigating that space of sounds. The range of sounds is defined by a dataset provided by the designer, while the means of navigation is defined by a combination of data labels and the selection of a sub-manifold from the latent space learned by the GAN. Our proposed system takes advantage of the rich latent space of GAN that consists of sounds that fill out the spaces “between” real data-like sounds. This augmented data from GAN is then used to train an RNN for its ability to respond immediately and continuously to parameter changes and to generate audio over unlimited periods of time. Furthermore, we develop a self organizing map technique for "smoothing" the latent space of GAN that results in perceptually smooth interpolation between audio timbres. We validate this process through user studies. The system contributes advances to the state of the art for generative sound model design that include system configuration and components for improving interpolation and the expansion of audio modeling capabilities beyond musical pitch and percussive instrument sounds into the more complex space of audio textures.

ABSTRACT. Guitar effects are commonly used in popular music to shape the guitar sound to fit specific genres or to create more variety within musical compositions. The sound is not only determined by the choice of the guitar effect, but also heavily depends on the parameter settings of the effect. Previous research focused on the classification of guitar effects and extraction of their parameter settings from solo guitar audio recordings. However, more realistic is the classification and extraction from instrument mixes. This work investigates the use of convolution neural networks (CNNs) for classification and extraction of guitar effects from audio samples containing guitar, bass, keyboard and drums. The CNN was compared to a baseline methods previously proposed like support vector machines and shallow neural networks together with predesigned features. The CNN outperformed all baselines, achieving a classification accuracy of up to 97.7\,\% and a mean absolute parameter extraction error of below 0.016 for the distortion, below 0.052 for the tremolo and below 0.038 for the slapback delay effect achieving or surpassing the presumed human expert error of 0.05.

ABSTRACT. There have been various generative music applications recently which employ a pre-trained transformer neural model. The way in which these models are trained effects greatly the nature of the music they produce, but there has been no exploration of the extent of this. We provide here a systematic evaluation of the output from GPT-2-based transformer models by analysing, comparing and contrasting the output from multiple models trained under various conditions, with reference to numerous musical metrics. We further describe a web application for exploring the generative output of such models and draw conclusions about how training effects output.

ABSTRACT. The latest language model by Open AI, GPT-3, has shown great potential for many NLP tasks, attracting attention from scientists and media alike, with applications in many different domains. In this work we carry out a first study on GPT-3’s capability to communicate musical decisions through textual explanations when prompted with a textual representation of a piece of music. Enabling a dialogue in human-AI music partnerships is an important step towards more engaging and creative human-AI interactions. Our results show that GPT-3 lacks the necessary intelligence to really understand musical decisions. A major barrier to reach a better performance is the lack of data that includes explanations of the creative process carried out by artists for musical pieces. We believe such a resource would aid the understanding and collaboration with AI music systems.

ABSTRACT. The evolutionary multi-objective algorithms have been widely applied for clustering. However, in general, the detection of heterogeneous nested clusters remains challenging for clustering algorithms. This paper proposes an adaptation of the connectedness criterion used as an objective function in established Evolutionary Multi-Objective Clustering approaches (EMOCs). This adaptation can improve the conflict between the objective functions, and then it promotes the detection of nested clusters. We performed experiments with four EMOCs (MOCK, MOCLE, $\Delta$-MOCK, and EMO-KC) that provide different features. These different EMOCs have different initialization methods and representations schemes, allowing us to analyze how the proposed objective function can contribute to detecting nested clusters. Our results show that our adapted objective function promotes a general gain in the performance of all these algorithms.

ABSTRACT. The evolution of hardware has enabled Artificial Neural Networks to become a staple solution to many modern Artificial Intelligence problems such as natural language processing and computer vision. The neural network's effectiveness is highly dependent on the optimizer used during training, which motivated significant research into the design of neural network optimizers. Current research focuses on creating optimizers that perform well across different topologies and network types. While there is evidence that it is desirable to fine-tune optimizer parameters for specific networks, the benefits of designing optimizers specialized for single networks remain mostly unexplored.

In this paper, we propose an evolutionary framework called Adaptive AutoLR (ALR) to evolve adaptive optimizers for specific neural networks in an image classification task. The evolved optimizers are then compared with state-of-the-art, human-made optimizers on two popular image classification problems. The results show that some evolved optimizers perform competitively in both tasks, even achieving the best average test accuracy in one dataset. An analysis of the best evolved optimizer also reveals that it functions differently from human-made approaches. The results suggest ALR can evolve novel, high-quality optimizers motivating further research and applications of the framework.

ABSTRACT. We present SLUG, a method that uses genetic algorithms as a wrapper for genetic programming (GP), to perform feature selection while inducing models. This method is first tested on four regular binary classification datasets, and then on 10 synthetic datasets produced by GAMETES, a tool for embedding epistatic gene-gene interactions into noisy datasets. We compare the results of SLUG with the ones obtained by other GP-based methods that had already been used on the GAMETES problems, concluding that the proposed approach is very successful, particularly on the epistatic datasets. We discuss the merits and weaknesses of SLUG and its various parts, i.e. the wrapper and the learner, and we perform additional experiments, aimed at comparing SLUG with other state-of-the-art learners, like decision trees, random forests and extreme gradient boosting. Despite the fact that SLUG is not the most efficient method in terms of training time, it is confirmed as the most effective method in terms of accuracy.

ABSTRACT. This paper focuses on the covering mechanism which generates a new if-then rule when the input data does not match the rules in the XCS Classifier System (XCS), a rule-based machine learning system, and discusses how the new rule should be generated from the viewpoint of “inheritance” and “expansion” of the generalization degree of the nearest neighbor rule in the continuous space. For this purpose, this paper proposes the two covering mechanisms based on the “inheritance” and “expansion” of the generalization degree of the nearest neighbor rule and compares their results by applying them to XCS with Continuous-valued inputs (XCSR). Through the intensive experiments on the three types of problems with the different characteristics, the following implications have been revealed: (1) the new rules should be generated by inheriting the generalization degree of the nearest neighbor rule in comparison with expanding it in the continuous space; and (2) XCSR with the “inheritance” based covering mechanism achieves higher classification accuracy with fewer rules than the conventional XCSR, which achieves higher classification accuracy than XCSR with the “expansion” based covering mechanism.

ABSTRACT. Given that physics-based models can be difficult to derive, data-driven models have been widely used for remaining useful life (RUL) prediction, which is a key element for predictive maintenance. In industrial applications, although the models have to be trained in a short time with limited computational resources, recent research using back propagation neural networks (BPNNs) has focused only on minimizing the RUL prediction error, without considering the time needed for training. Driven by this motivation, here we consider a simple and fast neural network, named extreme learning machine (ELM), and we optimize it for the specific case of RUL prediction. In particular, we propose to apply both single-objective and multi-objective optimization to search for the best ELM architectures in terms of a trade-off between RUL prediction error and training time, the latter being determined by the number of trainable parameters. We perform a comparative analysis on a recent benchmark dataset, the N-CMAPSS, in which we compare the proposed methods with other algorithms based on BPNNs. The results show that while the optimized ELMs perform slightly worse than the BPNNs in terms of RUL prediction error, they require a significantly shorter (up to 2 orders of magnitude) training time.

ABSTRACT. Granular metamaterials are a promising choice as a physical substrate for the realization of mechanical computing devices. In this paper, we demonstrate the design of two logic gates (an AND gate and an XOR gate) embedded in a granular metamaterial. We investigate two different setups, one with the mass-contrasting particles and the other with the stiffness-contrasting particles. We use Evolutionary Algorithms to design the placement of the particles to achieve a desired functionality. Our simulations confirm the existence of gradients with respect to the defined AND-ness and XOR-ness metrics in the search space. We verify the functionality of the best designed gates by probing the response of the system in the frequency space. Moreover, our results show the advantage of a stiffness-contrasting setup in providing more distinguishable bit representations at the output. We believe that our work enables a new generation of programmable acoustic logic gates that are embedded in material.

ABSTRACT. Understanding the search dynamics of multiobjective evolutionary algorithms (MOEAs) is still an open problem. This paper extends a recent network-based tool, search trajectory networks (STNs), to model the behavior of MOEAs. Our approach uses the idea of decomposition, where a multiobjective problem is transformed into several single-objective problems. We show that STNs can be used to model and distinguish the search behavior of two popular multiobjective algorithms, MOEA/D and NSGA-II, using 10 continuous benchmark problems with 2 and 3 objectives. Our findings suggest that we can improve our understanding of MOEAs using STNs for algorithm analysis.

ABSTRACT. This article presents a multiobjective variation of the problem of locating electric vehicle charging stations (EVCS) in a city known as the Multiobjective Electric Vehicle Charging Stations Locations (MO-EVCS-L) problem. MO-EVCS-L considers two conflicting objectives: maximizing the quality of service of the charging station network and minimizing the deployment cost when installing different types of charging stations. Two multiobjective metaheuristics are proposed to address MO-EVCS-L: the Non-dominated Sorting Genetic Algorithm, version II (NSGA-II) and the Strength Pareto Evolutionary Algorithm, version 2(SPEA2). The experimental analysis is performed on a real-world case study defined in Malaga, Spain, and it compares the proposed approaches with a baseline algorithm. Results show that the SPEA2 computes the most competitive solutions. Even though both metaheuristics found an accurate set of solutions that provide different tradeoffs between the quality of service and the installation costs.

ABSTRACT. Previous research on music related generation and transformation has commonly targeted single instrument or single melody music. Here, in contrast, five music genres are used with the goal to achieve selective remixing by using domain transfer methods on spectrogram images of music. A pipeline architecture comprised of two independent generative adversarial network models was created. The first, CycleGAN performs style transfer on constant-Q transform spectrogram images, by applying features from one of five genres to the spectrogram. The second network turns the spectrogram into a real-value tensor representation which is approximately reconstructed back into audio. Four seconds of music are output by the system and can be concatenated to recreate a full length music track. The system was evaluated through a number of experiments and a survey. Due to the increased complexity involved in processing high sample rate music with homophonic or polyphonic audio textures, the system’s audio output was considered to be low quality, but the style transfer produced noticeable selective remixing on most of the music tracks used for evaluation.

ABSTRACT. The field of automatic music composition has seen great progress in recent years, specifically with the invention of transformer-based architectures. When using any deep learning model which considers music as a sequence of events with multiple complex dependencies, the selection of a proper data representation is crucial. In this paper, we tackle the task of conditional drums generation using a novel data encoding scheme inspired by the Compound Word representation, a tokenization process of sequential data. Therefore, we present a sequence-to-sequence architecture where a Bidirectional Long short-term memory (BiLSTM) Encoder receives information about the conditioning parameters (i.e., accompanying tracks and musical attributes), while a Transformer-based Decoder with relative global attention produces the generated drum sequences. We conducted experiments to thoroughly compare the effectiveness of our method to several baselines. Quantitative evaluation shows that our model is able to generate drums sequences that have similar statistical distributions and characteristics to the training corpus. These features include syncopation, compression ratio, and symmetry among others. We also verified, through a listening test, that generated drum sequences sound pleasant, natural and coherent while they “groove” with the given accompaniment.

ABSTRACT. Scramble is a standalone MIDI tool developed in Max/MSP for the real-time generation of polyphonic music, that combines Markov chains and LSTM neural networks. It offers a simplified user interface for the analysis of MIDI files, the creation of models including expressive parameters beside pitch and rhythm, and the interactive control of the generated output. In this paper we describe and motivate the strategies we implemented for the analysis and representation of the data, and the encoding techniques we resorted to in order to facilitate the detection of low-level relationships whilst saving computational resources. We also propose a representation of pitch and time domains suitable for both the Markov chain and the LSTM modules and detail the tool’s architecture from a functional standpoint and from the perspective of the user. We conclude by presenting the testing results and the future developments of the system.

ABSTRACT. The use of computer generated characters as artificial dancers offers interesting creative possibilities, especially when endowing these characters with morphologies and behaviours that differ significantly from those of human dancers. At the same time, it is challenging to create movements for such non-anthropomorphic characters that are at the same time expressive and physically plausible. Motion synthesis techniques based either on data driven methods or physics simulation each have their own limitations concerning the aspects of movements and the range of morphologies they can be used for. This paper presents a proof of concept system that combines a data driven method with a physics simulation for the purpose of synthesizing expressive movements for computer generated characters with arbitrary morphologies. A core component of the system is a reinforcement learning algorithm that employs reward functions based on Laban Effort Factors. This system has been tested by training three different non-anthropomorphic morphologies on different combinations of these reward functions. The results obtained so far clearly indicate that the system is able to generate a large diversity of poses and motions which specifically reflect the characteristics of each morphology and Effort Factor.

ABSTRACT. Membrane computing is a discipline that aims to perform computation by mimicking nature at the cellular level. Spiking Neural P (in short, SN P) systems are a subset of membrane computing methodologies that combine spiking neurons with membrane computing techniques, where "P" means that the system is intrinsically parallel. While these methodologies are very powerful, being able to simulate a Turing machine with only few neurons, their design is time-consuming and it can only be handled by experts in the field, that have an in-depth knowledge of such systems. In this work, we use the Neuroevolution of Augmenting Topologies (NEAT) algorithm, usually employed to evolve multilayer perceptrons and recurrent neural networks, to evolve SN P systems. Unlike existing approaches for the automatic design of SN P systems, NEAT provides high flexibility in the type of SN P systems, removing the need to specify a great part of the system. To test the proposed method, we evolve Spiking Neural P systems as policies for two classic control tasks from OpenAI Gym. The experimental results show that our method is able to generate efficient (yet extremely simple) Spiking Neural P systems that can solve the two tasks. A further analysis shows that the evolved systems act on the environment by performing a kind of "if-then-else" reasoning.

ABSTRACT. Using Neuroevolution combined with Novelty Search to promote behavioural diversity is capable of constructing high-performing ensembles for classification. However, using gradient descent to train evolved architectures during the search can be computationally prohibitive. Here we propose a method to overcome this limitation by using a surrogate model which estimates the behavioural distance between two neural network architectures required to calculate the sparseness term in Novelty Search. We demonstrate a speedup of 10 times over previous work and significantly improve on previous reported results on three benchmark datasets from Computer Vision --- CIFAR-10, CIFAR-100, and SVHN. This results from the expanded architecture search space facilitated by using a surrogate. Our method represents an improved paradigm for implementing horizontal scaling of learning algorithms by making an explicit search for diversity considerably more tractable for the same bounded resources.

ABSTRACT. On-line learning mechanisms are frequently employed to implement self-adaptivity in modern systems and make them independent from design-time assumptions. With more widespread use in technical systems that interact with their physical environment, e.g. cyber-physical systems, the fulfillment of safety requirements is increasingly gaining attention. We focus on the learning classifier system XCS with its human-interpretable rules and propose an approach to integrate safety guarantees directly into its rule base. We leverage the interpretability of XCS' rules to internalize the safety-critical knowledge, as opposed to related work, which often relies on an external safety monitor. The experimental evaluation shows that such manually injected knowledge not only gives safety guarantees but aids the learning mechanism of XCS. Especially in complex environments where XCS is struggling to find the optimal solution, the use of forbidden classifiers leads to a considerably better performance than with an external safety monitor.

ABSTRACT. In software development, third-party libraries are commonly used to reduce implementation efforts and errors, while delivering high-quality, reliable and secure software. To support software evolution, newer libraries are continuously released to offer added features, and critical updates such as bug and vulnerability fixes. As a result, old (source) libraries and their methods must be replaced with their newer, updated counterparts (target libraries) during the library migration process. This is known to be a time-consuming and error-prone process as developers need to analyze both the source and target library's Application Programming Interface (API) documentation and implementation to replace every source API with target API(s).

Recent studies have utilized various techniques to recommend the appropriate target library for replacement, but do not provide guidelines on how to replace each source API with its target library APIs. To address this limitation, our work leverages evolutionary search algorithms to recommend at the method-level by (1) formulating API migration as a combinatorial optimization problem, and (2) using genetic algorithms (GA) to recommend suitable APIs during migration based on the method signature and documentation similarity and co-occurrence. We conduct an empirical study on 9 popular library migrations from 57,447 open-source Java projects and demonstrate that GA can recommend multiple APIs for replacement with up to 100% precision for certain library migrations.

ABSTRACT. Even when deep convolutional neural networks have proven to be powerful at saliency detection, they have a vulnerability that should not be ignored: they are susceptible to adversarial attacks, which makes them highly unreliable. Reliability is an important aspect to consider when it comes to salient object detection; without it, any attacker can render the algorithm useless. Brain programming is an evolutionary methodology for visual problems that is highly resilient, and is able to withstand even the most intense perturbations. In this work, we perform for the first time a study that compares the resilience against adversarial attacks and noise perturbations using a real-world database of a shorebird called the Snowy Plover in a visual attention task. The images in this database were taken on the field and even pose a detection challenge due to the nature of the environment and the physical characteristics of the bird. By attacking three different deep convolutional neural networks using adversarial examples from this database, we prove that they are no match for the brain programming algorithm when it comes to resilience, suffering great losses in their performance. On the other hand, brain programming stands its ground and sees its performance unaffected. Also, by using images of the Snowy Plover, we refer to the importance of resilience in real-world issues where conservation is present. Brain programming is the first highly resilient evolutionary algorithm used for saliency detection tasks.

ABSTRACT. Combinatorial designs provide an interesting source of optimization problems. Among them, permutation codes are particularly interesting given their applications in powerline communications, flash memories, and block ciphers. This paper addresses the design of permutation codes by evolutionary algorithms (EA) by developing an iterative approach. Starting from a single random permutation, new permutations satisfying the minimum distance constraint are incrementally added to the code by using a permutation-based EA. We investigate our approach against four different fitness functions targeting the minimum distance requirement at different levels of detail and with two different policies concerning code expansion and pruning. We compare the results achieved by our EA approach to those of a simple random search, remarking that neither method scales well with the problem size.