ABSTRACT. Genetic Programming (GP) based symbolic regression is prone to generate complex models, which often overfit the training data and generalise poorly onto unseen data. To address this issue, many pieces of research have been devoted to controlling the model complexity of GP. One recent work aims to control model complexity using a new representation called Adaptive Weighted Splines. With its semi-structured characteristic, the Adaptive Weighted Splines representation can control the model complexity explicitly, which was demonstrated to be significantly better than its tree-based counterpart at generalising onto unseen data. This work seeks to significantly extend the previous work by proposing a multi-objective GP algorithm with the Adaptive Weighted Splines representation, which utilises parsimony pressure to further control the model complexity, as well as improve the interpretability of the learnt models. Experiment results show that, compared with single-objective GP with the Adaptive Weighted Splines and multi-objective tree-based GP with parsimony pressure, the new multi-objective GP method generally obtains superior fronts and produces better generalising models. These models are also significantly smaller and more interpretable.

ABSTRACT. Development is fundamental for living beings. As robots are often designed to mimic biological organisms, development is believed to be crucial for achieving successful results in robotic agents, as well. What is not clear, though, is the most appropriate scheduling for development. While in real life systems development happens mostly during the initial growth phase of organisms, it has not yet been investigated whether such assumption holds also for artificial creatures. In this paper, we employ a evolutionary approach to optimize the development---according to different representations---of Voxel-based Soft Robots (VSRs), a kind of modular robots. In our study, development consists in the addition of new voxels to the VSR, at fixed time instants, depending on the development schedule. We experiment with different schedules and show that, similarly to living organisms, artificial agents benefit from development occurring at early stages of life more than from development lasting for their entire life.

ABSTRACT. Evolutionary design of reduced precision classifiers of levodopa-induced dyskinesia (LID) is investigated in order to find a hardware-efficient classifier together with classification accuracy as close as possible to a baseline software implementation. Cartesian genetic programming (CGP) equipped with the coevolution of adaptive size fitness predictors (coASFP) is used to design LID-classifiers working with fixed-point arithmetics with reduced precision, which is suitable for implementation in application-specific integrated circuits. In this particular task, we achieved a significant evolutionary design computational cost reduction in comparison with the original CGP. Moreover, coASFP effectively prevented overfitting in this task. Experiments with reduced precision LID-classifier design shown that evolved classifiers working with 8-bit unsigned integer data representation, together with the input data scaling using the logical right shift, not only significantly outperformed hardware characteristics of all other investigated solutions but also achieved a better classifier accuracy in comparison with classifiers working with the floating-point numbers.

ABSTRACT. Human code is different from code generated by program search. We investigate if properties from human-generated code can guide program search to improve the qualities of the generated programs, e.g., readability and performance. Here we focus on program search with grammatical evolution, which produces code that has different structure compared to human-generated code, e.g., loops and conditions are hardly used. We use a large code-corpus that was mined from the open software repository service GitHub and measure software metrics and properties describing the code-base. We use this knowledge to guide the search by incorporating a new selection scheme. Our new selection scheme favors programs that are structurally similar to the programs in the GitHub code-base. We find noticeable evidence that software metrics can help in guiding evolutionary search.

ABSTRACT. Anticipatory Learning Classifier Systems (ALCS) are rule-based machine learning algorithms that can simultaneously develop a complete representation of their environment and a decision policy based on this representation to solve their learning tasks. This paper introduces BEACS (Behavioral Enhanced Anticipatory Classifier System) in order to handle non-deterministic partially observable environments and to allow users to better understand the environmental representations issued by the system. BEACS is an ALCS that enhances and merges Probability-Enhanced Predictions and Behavioral Sequences approaches used in ALCS to handle such environments. The Probability-Enhanced Predictions consist in enabling the anticipation of several states, while the Behavioral Sequences permits the construction of sequences of actions. The capabilities of BEACS have been studied on a thorough benchmark of 23 mazes and the results show that BEACS can handle different kinds of non-determinism in partially observable environments, while describing completely and more accurately such environments. BEACS thus provides explanatory insights about created decision polices and environmental representations.

ABSTRACT. Program synthesis automates the process of writing code, which can be a very useful tool in allowing people to better leverage computational resources. However, a limiting factor in the scalability of current program synthesis techniques is the large size of the search space, especially for complex programs. We present a new model for synthesizing programs which reduces the search space by composing programs from program pieces, which are component functions provided by the user. Our method uses genetic programming search with a fitness function based on refinement type checking, which is a formal verification method that checks function behavior expressed through types. We evaluate our implementation of this method on a set of 3 benchmark problems, observing that our fitness function is able to find solutions in fewer generations than a fitness function that uses example test cases. These results indicate that using refinement types and other formal methods within genetic programming can improve the performance and practicality of program synthesis.

ABSTRACT. This work presents the first study of using the popular Monte Carlo Tree Search (MCTS) method combined with dedicated heuristics for solving the weighted vertex coloring problem. Starting with the basic MCTS algorithm, we gradually introduce a number of algorithmic variants where MCTS is extended by various simulation strategies including greedy and local search heuristics. We conduct experiments on well-known benchmark instances to assess the value of each studied combination. We also provide empirical evidence to shed light on the advantages and limits of each strategy.

ABSTRACT. This work utilizes Algorithm Selection for solving the Team Orienteering Problem (TOP). The TOP is an NP-hard combinatorial optimization problem in the routing domain. This problem has been modelled with various extensions to address different real-world problems like tourist trip planning. The complexity of the problem motivated to devise new algorithms. However, none of the existing algorithms came with the best performance across all the widely used benchmark instances. This fact suggests that there is a performance gap to fill. This gap can be targeted by developing more new algorithms as attempted by many researchers before. An alternative strategy is performing Algorithm Selection that will automatically choose the most appropriate algorithm for a given problem instance. This study considers the existing algorithms for the Team Orienteering Problem as the candidate method set. For matching the best algorithm with each problem instance, the specific instance characteristics are used as the instance features. An algorithm Selection approach, namely ALORS, is used to conduct the selection mission. The computational analysis based on 157 instances showed that Algorithm Selection outperforms the state-of-the-art algorithms despite the simplicity of the Algorithm Selection setting. Further analysis illustrates the match between certain algorithms and certain instances. Additional analysis showed that the time budget significantly affects the algorithms' performance.

ABSTRACT. Designing efficient heuristics is a laborious and tedious task that generally requires a full understanding and knowledge of a given optimization problem. Hyper-heuristics have been recently introduced to tackle this issue and are mostly relying on Genetic Programming and its variants. Many attempts in the literature have shown that an automatic training mechanism for heuristic learning is possible and can challenge human-based heuristics in terms of gap to optimality. In this work, we introduce a novel approach based on a recent work on Deep Symbolic Regression. We demonstrate that scoring functions can be trained using Recurrent Neural Networks to tackle a well-know combinatorial problem, i.e., the Multi-dimensional Knapsack. Experiments have been conducted on instances from the OR-Library and results show that the proposed modus operandi outperforms state-of-the-art approaches including human-based heuristics and classical heuristic generation approaches.

ABSTRACT. In this work, we present the benchmark generator PUBOi, Polynomial Unconstrained Binary Optimization, that combines subproblems to create instances of pseudo-boolean optimization problems. This Walsh basis of functions makes it possible to represent any mono-objective pseudo-Boolean functions including existing classical optimization problems. The benchmark generator can tune main features of problems such as problem dimension, non-linearity degree, neutrality. Additionally, to be able create instances with properties similar to those of real-like combinatorial optimization problems, the goal of PUBOi is to introduce the notion of variable importance. Indeed, the importance of decision variables can be tuned using three benchmark parameters. In the version presented here, we consider four subproblems as a basis already used in Chook generator for benchmarking quantum computers, and algorithms. We also present the impact of benchmark parameters using a fitness landscape analysis that empirically shows these parameters significantly impact the variable importance.

ABSTRACT. Natural Evolution Strategies (NES) is a promising framework for black-box continuous optimization problems. NES optimizes the parameters of a probability distribution based on the estimated natural gradient, and one of the key parameters affecting the performance is the learning rate. We argue that from the viewpoint of the natural gradient method, the learning rate should be determined according to the estimation accuracy of the natural gradient. To do so, we propose a new learning rate adaptation mechanism for NES. The proposed mechanism makes it possible to set a high learning rate for problems that are relatively easy to optimize, which results in speeding up the search. On the other hand, in problems that are difficult to optimize (e.g., multimodal functions), the proposed mechanism makes it possible to set a conservative learning rate when the estimation accuracy of the natural gradient seems to be low, which results in the robust and stable search. The experimental evaluations on unimodal and multimodal functions demonstrate that the proposed mechanism works properly depending on a search situation and is effective over the existing method, i.e., using the fixed learning rate.

ABSTRACT. The choice of parameter values in evolutionary algorithms greatly affects their performance. Many popular parameter tuning techniques are limited by the tuning budget for finding a good set of parameter values. Fitness landscape analysis can show similarities between different problem instances. Reusing the performance data from similar problem instances may help improve the parameter tuning approaches and reduce their tuning budget requirements.

In this paper, we present an approach to automated parameter tuning for the (1 + (λ, λ)) genetic algorithm using fitness landscape analysis and machine learning. We collected a dataset of good parameter choices for multiple instances of the W-model problem with landscape features computed in flacco. Using a neural network trained on this data we are able to suggest good parameter values for the algorithm on different instances of the W-model problem, as well as linear integer weights problem and MAX-3SAT problem.

ABSTRACT. A network-based modelling technique, search trajectory networks (STNs), has recently helped to understand the dynamics of neuroevolution algorithms such as NEAT. Modelling and visualising variants of NEAT made it possible to analyse the dynamics of search operators. Thus far, this analysis was applied directly to the NEAT genotype space composed of neural network topologies and weights. Here, we extend this work, by illuminating instead the behavioural space, which is available when the evolved neural networks control the behaviour of agents. Recent interest in \textit{behaviour characterisation} highlights the need for divergent search strategies. Quality-diversity and Novelty search are examples of divergent search, but their dynamics are not yet well understood. In this article, we examine the idiosyncrasies of three neuroevolution variants: novelty, random and objective search operating as usual on the genotypic search space, but analysed in the behavioural space. Results show that novelty is a successful divergent search strategy. However, its abilities to produce diverse solutions are not always consistent. Our visual analysis highlights interesting relationships between topological complexity and behavioural diversity which may pave the way for new characterisations and search strategies.

ABSTRACT. In this paper we study anisotropic consensus-based optimization (CBO), a multi-agent metaheuristic derivative-free optimization method capable of globally minimizing nonconvex and nonsmooth functions in high dimensions. CBO is based on stochastic swarm intelligence, and inspired by consensus dynamics and opinion formation. Compared to other metaheuristic algorithms like particle swarm optimization, CBO is of a simpler nature and therefore more amenable to theoretical analysis. By adapting a recently established proof technique, we show that anisotropic CBO converges globally with a dimension-independent rate for a rich class of objective functions under minimal assumptions on the initialization of the method. Moreover, the proof technique reveals that CBO performs a convexification of the optimization problem as the number of agents goes to infinity, thus providing an insight into the internal CBO mechanisms responsible for the success of the method. To motivate anisotropic CBO from a practical perspective, we further test the method on a complicated high-dimensional benchmark problem, which is well understood in the machine learning literature.

ABSTRACT. Dynamic job shop scheduling has a wide range of applications in reality such as order picking in warehouse. Using genetic programming to design scheduling heuristics for dynamic job shop scheduling problems becomes increasingly common. In recent years, multitask genetic programming-based hyper-heuristic methods have been developed to solve similar dynamic scheduling problem scenarios simultaneously. However, all of the existing studies focus on the tree-based genetic programming. In this paper, we investigate the use of linear genetic programming, which has some advantages over tree-based genetic programming in designing multitask methods, such as building block reusing. Specifically, this paper makes a preliminary investigation on several issues of multitask linear genetic programming. The experiments show that the linear genetic programming within multitask frameworks have a significantly better performance than solving tasks separately, by sharing useful building blocks.

ABSTRACT. Denoising Autoencoder Genetic Programming (DAE-GP) is a novel neural network-based estimation of distribution genetic programming (EDA-GP) algorithm that uses denoising autoencoder long short-term memory networks as a probabilistic model to replace the standard mutation and recombination operators of genetic programming (GP). At each generation, the idea is to flexibly identify promising properties of the parent population and to transfer these properties to the offspring where the DAE-GP uses denoising to make the model robust to noise that is present in the parent population. Denoising partially corrupts candidate solutions that are used as input to the model. The stronger the corruption, the stronger the generalization of the model.In this work, we study how corruption strength affects the exploration and exploitation behavior of the DAE-GP. For a generalization of the royal tree problem (high-locality problem), we find that the stronger the corruption, the stronger the exploration of the solution space. For the given problem, weak corruption resulting in a stronger exploitation of the solution space performs best. However, in more rugged fitness landscapes (low-locality problems), we expect that a stronger corruption resulting in a stronger exploration will be helpful. Choosing the right denoising strategy can therefore help to control the exploration and exploitation behavior in search, leading to an improved search quality.

ABSTRACT. In this paper, we are interested in ad hoc autonomous agent team composition using cooperative co-evolutionary algorithms (CCEA). In order to accurately capture the individual contribution of team agents, we propose to limit the number of agents which are updated in-between team evaluations. However, this raises two important problems with respect to (1) the cost of accurately estimating the marginal contribution of agents with respect to the team learning speed and (2) completing tasks where improving team performance requires multiple agents to update their policies in a synchronized manner. We introduce a CCEA algorithm that is capable of learning how to update just the right amount of agents' policies for the task at hand. We use a variation of the El Farol Bar problem, formulated as a multi-robot resource selection problem, to provide an experimental validation of the algorithms proposed.

ABSTRACT. The real-time Train Routing Selection Problem (rtTRSP) is the combinatorial optimization problem of selecting, for each train in a rail network, the best routing alternatives. Solving the rtTRSP aims to limit the search space of train rescheduling problems, highly affected by the number of routing variables. The rtTRSP is modelled as the minimum weight clique problem in an undirected k-partite graph. This problem is NP-hard and its complexity is highly increased by the size of the instances and the short available computation time. A sequential version of the Ant Colony Optimization (ACO) for the rtTRSP has been proposed in the literature. However, in large instances the algorithm struggles to find high-quality solutions. This paper proposes the implementation and the performance evaluation of a parallel ACO for large rtTRSP instances. Specifically, we analyze the performance of the algorithm by standard parallel metrics. In addition, we test two parallel local search strategies, which consider different solution neighbourhoods. Computational experiments are performed on the practical case study of Lille Flandres station area in France. The results show a significant speed-up of the rtTRSP search space exploration and more promising diversification patterns. Both these improvements enhance the rtTRSP solutions.

ABSTRACT. The administration of access control structures in Enterprise Resource Planning Systems (ERP) is mainly organized by Role Based Access Control. The associated optimization problem is called the Role Mining Problem (RMP), which is known to be NP-complete. The goal is to search for role concepts minimizing the number of roles. Algorithms for this task are presented in literature, but often they cannot be used for role mining in ERP in a straightforward way, as ERP systems have some additional conditions and constraints. Some ERP systems require multiple levels of roles. This paper presents three approaches to computing such hierarchical role concepts.

One is aiming at optimizing multiple levels of roles simultaneously. The other approaches divide the multi-level role mining problem into separate sub-problems, which are optimized individually. All approaches are based on an evolutionary algorithm for single-level role mining and have been implemented and evaluated in a range of experiments.

ABSTRACT. Evolving dispatching rules through Genetic Programming (GP) has been shown to be successful for solving Dynamic Job Shop Scheduling (DJSS). However, the GP-evolved rules are often very large and complex, and are hard to interpret. Simplification is a promising technique that can reduce the size of GP individuals without sacrificing effectiveness. However, GP with simplification has not been studied in the context of evolving DJSS rules. This paper proposes a new GP with simplification for DJSS, and analyses its performance in evolving both effective and simple/small rules. In addition to adopting the generic algebraic simplification operators, we also developed new problem-specific numerical and behavioural simplification operators for DJSS. The results show that the proposed approach can obtain better and simpler rules than the baseline GP and existing GP algorithms with simplification. Further analysis verified the effectiveness of the newly developed numerical and simplification operators.

ABSTRACT. Topology optimisation of trusses can be formulated as a combinatorial and multi-modal problem in which locating distinct optimal designs allows practitioners to choose the best design based on their preferences. Bilevel optimisation has been successfully applied to truss optimisation to consider topology and sizing in upper and lower levels, respectively. We introduce exact enumeration to rigorously analyse the topology search space and remove randomness for small problems. We also propose novelty-driven binary particle swarm optimisation for bigger problems to discover new designs at the upper level by maximising novelty. For the lower level, we employ a reliable evolutionary optimiser to tackle the layout configuration aspect of the problem. We consider truss optimisation problem instances where designers need to select the size of bars from a discrete set with respect to practice code constraints. Our experimental investigations show that our approach outperforms the current state-of-the-art methods and it obtains multiple high-quality solutions.

ABSTRACT. The performance of cloud computing depends in part on job-scheduling algorithms, but also on the connection structure. Previous work on this structure has mostly looked at fixed and static connections. However, we argue that such static structures cannot be optimal in all situations. We introduce a dynamic hierarchical connection system of sub-schedulers between the scheduler and servers, and use artificial intelligence search algorithms to optimise this structure. Due to its dynamic and flexible nature, this design enables the system to adaptively accommodate heterogeneous jobs and resources to make the most use of resources. Experimental results compare genetic algorithms and simulating annealing for optimising the structure, and demonstrate that a dynamic hierarchical structure can significantly reduce the total makespan (max processing time for given jobs) of the heterogeneous tasks allocated to heterogeneous resources, compared with a one-layer structure. This reduction is particularly pronounced when resources are scarce.

ABSTRACT. Catan is a popular multiplayer board game that involves multiple gameplay notions: stochastic elements related to the dice rolls as well as to the the initial placement of resources on the map and the drawing of development cards, strategic notions for the placement of the cities and the roads which call upon topological and shape recognition notions and notions of expectation of gains linked to the probabilities of the rolls of the dice. In this paper, we develop a policy for this game using a convolutional neural network. The used deep reinforcement learning algorithm is Expert Iteration \cite{anthony2017thinking} which has already given excellent results for Alpha Zero and its descendants.

ABSTRACT. Speedrunning in general means to play a video game fast, i.e. using all means at one's disposal to achieve a given goal in the least amount of time possible. To do so, a speedrun must be planned in advance, or routed, as referred to by the community. This paper focuses on discovering challenges and defining models needed when trying to approach the problem of routing algorithmically. To do so, this paper is split in two parts. The first part provides an overview of relevant speedrunning literature, extracting vital information and formulating criticism. Important categorizations are pointed out and a nomenclature is build to support professional discussion. The second part of this paper then refers to the actual speedrun routing optimization problem. Different concepts of graph representations are presented and their potential is discussed. Visions both for problem modeling as well as solving are presented and assessed regarding suitability and expected challenges. Finally, a first assessment of the applicability of existing optimization methods to the defined problem is made, including metaheuristics/EA and Deep Learning methods.

ABSTRACT. Resilience can be defined as a system's capability for returning to normal operation after having suffered a disruption. This notion is of the foremost interest in many areas, in particular engineering. We argue in this position paper that is is a crucial property for bioinspired optimization algorithms as well. Following a computer system perspective, we correlate some of the defining requirements for attaining resilient systems to issues, features, and mechanisms of these techniques. It is shown that bioinspired algorithms do not only exhibit a notorious built-in resilience, but that their plasticity also allows accommodating components that may boost it in different ways. We also provide some relevant research directions in this area.

ABSTRACT. This paper proposes an enhanced feature selection (FS) approach to improve the classification tasks, taking into account data dimensionality as a significant criterion of the dataset. High dimensionality may cause serious problems in classification that degrade the performance of the classifier. Among these problems: generating complex models (overfitting), increasing the learning time, and including redundant and irrelevant features in the learning model. FS is a data mining technique to minimize the number of dimensions (features) by getting rid of redundant and irrelevant features. Meanwhile, FS tries to maximize the classification performance. As FS is an optimization problem, meta-heuristic optimization algorithms can take place to achieve superior results in solving such problems. This paper proposes the Moth Flame Optimization (MFO) algorithm to tackle the FS problem. A new initialization method called opposition-based is proposed. Furthermore, a new update strategy is proposed to alleviate the local minima. The comparative results find that the proposed approach improves the MFO performance and outperforms other similar approaches.

ABSTRACT. Predicting the performance of an optimization algorithm on a new problem instance is crucial in order to select the most appropriate algorithm for solving that problem instance. For this purpose, recent studies learn a supervised machine learning (ML) model using a set of problem landscape features linked to the performance achieved by the optimization algorithm. However, these models are black-box with the only goal of achieving good predictive performance, without providing explanations which landscape features contribute the most to the prediction of the performance achieved by the optimization algorithm. In this study, we investigate the expressiveness of problem landscape features utilized by different supervised ML models in automated algorithm performance prediction. The experimental results point out that we need to treat the automated algorithm performance prediction as a personalized task that requires training of supervised ML models using different problem landscape feature portfolios depending on which problem instance is solved. We have shown that the selection of the supervised ML method is crucial, since different supervised ML regression models utilize the problem landscape features differently and there is no common pattern with regard to which landscape features are the most informative.

ABSTRACT. Using a well known benchmarking and profiling environment, we compare the performances of three simple and easy to use metaheuristics for global optimization: Differential Evolution, Basin Hopping and Particle Swarm Optimization. The comparison was done on a test set of 24 functions featuring many characteristics found on real-world problems and on four different space dimensions. Our results statistically show that there is no clear winner overall. The three methods perform well in general and the actual differences are related to the different groups of functions in the benchmark with Basin Hopping being the most robust technique, and Differential Evolution and Particle Swarm Optimization excelling on highly multi-modal functions.

ABSTRACT. Using differential evolution and statistical analysis, this paper investigates a methodology that is capable of determining the ion channels in a neuron from membrane potential data obtained by the current-clamp method. These data provide the aggregated electrical response of the neuron under stimulation by integrating the individual responses of the different ion channels involved. The proposed methodology aims at determining which are these ion channels based on the hypothesis that each ion channel provides a specific signature in the aggregated response that we are able to detect. In order to assess the methodology, we propose a benchmark of synthetic data where the types of ion channels are predefined in advance. Results show that the methodology is able to determine the correct ion channels in three out of the four data sets. Furthermore, we obtain some hints for future enhancements on the method.

ABSTRACT. Many creative generative design spaces contain multiple regions with individuals of high aesthetic value. Yet traditional evolutionary computing methods typically focus on optimisation, searching for the fittest individual in a population. In this paper we apply quality-diversity search methods to explore a creative generative system (an agent-based line drawing model). We perform a random sampling of genotype space and use individual artist-assigned evaluations of aesthetic quality to formulate a computable fitness measure specific to the artist and this system. To compute diversity we use a convolutional neural network to discriminate features that are dimensionally reduced into two dimensions. We show that the quality-diversity search is able to find multiple phenotypes of high aesthetic value. These phenotypes show greater diversity and quality than those the artist was able to find using manual search methods.

ABSTRACT. We present a novel music generation framework for music infilling which is the task of generating music sections given the surrounding multi-track music as input. The proposed transformer-based framework is extensible by adding new music control tokens such as tonal tension per bar and track polyphony level. We explore the effects of including several musically meaningful control tokens, and evaluate the results using pitch and rhythm objective metrics. Our results demonstrate that adding additional control tokens helps to generate music with stronger stylistic similarities to the original music. It also provides the user with more control to change the music texture and specific tonal tension per bar compared to previous research which only provided control for track density. We present the model in a Google Colab notebook to enable interactive generation.

ABSTRACT. This paper introduces a machine learning tool for generating novel furniture designs. We employ a network graph to represent object structure and utilise two deep neural nets in combination to learn these network graphs, reproduce them, and generate variations on them. We apply the tool to the domain of furniture design and describe how the tool can create novel designs quickly and in large volume. This original generative approach allows a designer to efficiently consider novel design candidates. The option to train the system on multiple product types allows designers to explore designs that live between and outside the traditional concept of the domain object. We suggest that this workflow could be generalised for use beyond the domain of furniture design.

ABSTRACT. We showcase visualizations created for art periods of Dalí, van Gogh, and Picasso by leveraging deep neural embedding models to represent color features. First, vector representations are generated for every color used in artworks of these painters. Next, t-SNE is applied to generate a two-dimensional visualizations of the color space. Colors used in close proximity on the canvas are observed as compact clusters in the visualizations. These visualizations are termed as fingerprints, as they uniquely depict each art period of a painter, by highlighting the color palette used in their works. The authors further provide commentary on the artists' art periods and how the fingerprints showcase their artistic evolution.

ABSTRACT. Since 2015, commercial gestural interfaces have widened accessibility for researchers and artists to use novel Electromyographic (EMG) biometric data. EMG data measures musclar amplitude and allows us to enhance Human-Computer Interaction (HCI) through providing natural gestural interaction with digital media. Virtual Reality (VR) is an immersive tool capable of simulating the real world and abstractions of it. However, current commercial VR technology is not equipped to process and use biometric information. Using biometrics within VR allows for better gestural detailing and use of complex custom gestures, such as those found within instrumental music performance, compared to using optical sensors for gesture recognition in current commercial VR equipment. Although, EMG data is complex and machine learning must be used to employ it. This study uses a Myo armband to classify four custom gestures in Wekinator and observe their prediction accuracies and representations (including or omitting signal onset) to compose music within VR. Results show that specific regression and classification models, according to gesture representation type, are the most accurate when classifying four music gestures for advanced music HCI in VR. We apply and record our results, showing that EMG biometrics are promising for future interactive music composition systems in VR.

ABSTRACT. Using computational processes to model aesthetic judgement is an important, if indispensable, endeavour when designing creative AI systems. In the music domain, studies have thus far concentrated on information-based approaches, albeit often offered as single-factor explanations, and developed in isolation from the ongoing debates taking place in other domains, particularly in cognitive science and philosophy of arts. In this paper, the notion of complexity is explored as a tool towards a cross-modal alternative to the methodological legacy rooted in isolated perceptual analyses of the aesthetic object. To this end, we consider and compare a set of metrics, some of which are long-established in this field (e.g., information dynamics), or derived/adapted from other domains (e.g., processing fluency theory, soundscape ecology), or originally designed. We apply these to the aesthetic evaluation of noise music in an effort to generalise our discourse to musical expressions that further foreground the limitations of established practices and the exciting road ahead for computational music aesthetics.

ABSTRACT. We analyze the bibliography related to portfolio optimization using metaheuristics and bioinspired algorithms. To this end, we perform data clustering based on lexical similarity between bibliographical descriptors and propose an internal arrangement of each cluster using evolutionary algorithms. We also conduct a network analysis in order to determine relevant keywords and their associations

ABSTRACT. Local search heuristics are commonly used for several classes of combinatorial problems, and neighbourhood functions play a critical role in their ability to adequately explore the search space. We suggest a statistical approach to the investigation of neighbourhood utility for specific algorithm-problem pairs. The approach is illustrated using a Simulated Annealing with six neighbourhood structures for the Unrelated Parallel Machine Scheduling Problem with Sequence Dependent Setup Times. The proposed approach may be useful for devising self-adaptation rules for the selection of neighbourhood functions conditionally on problem characteristics and the search stage.

ABSTRACT. Bio-inspired computation is implemented in a wide range of research fields. We propose the use of Evolutionary Algorithms (EA) to develop a path computation tool to send data over the network in the presence of Internet tunnels. We have chosen this algorithm for its evident success as a search and optimisation tool.

ABSTRACT. This is ongoing interdisciplinary research drawing inspiration from the rapidly growing field of evolutionary developmental biology, i.e., the so-called “Evo-Devo” paradigm. We not only aim to investigate the similarities but importantly, also the differences between the synthesis of a biological organism and an engineering system. A bespoke algorithm will be developed using an Evolutionary Algorithm to generate fractal patterns observed in nature. The success of this method will allow us to investigate its possible usage in designing an engineering system.

ABSTRACT. Artificial Spin Ice (ASI), meaning a collection of interacting nanomagnets, were originally constructed as a model system to study physical phenomena, but have become a promising avenue for in materio computation. Though the potential of ASI for computation is often cited within the ASI community, there are few concrete examples of it being used as such. Here we demonstrate evolving the geometry of an ASI such that it can perform a simple classification task. The success of the method on this simple problem gives confidence that this methodology could be used to create ASI capable of solving more complex tasks.

ABSTRACT. Multi-objective neuroevolution is a research field of growing importance within reinforcement learning. This paper introduces ANN-EMOA, a novel multi-objective neuroevolutionary algorithm that is inspired by nNEAT and aims at high efficiency, usability, and comprehensibility. To that end it applies a simple encoding and efficient variation operators. Diversity plays a key role in evolutionary computation. For this reason, we apply the Riesz s-energy to foster diversity explicitly. This paper also develops a new efficient approach to determine the individual Riesz s-energy contribution of each solution within a set. To assess the performance of the new ANN-EMOA it is compared to nNEAT and NEAT-MODS, two multi-objective variants of NEAT, in the multi-objective Double Pole Balancing problem. The results show that ANN-EMOA does not only converge faster and to higher quality-levels than its competitors, but it also maintains more compact network-genomes and shows convincing performance even with comparably small populations.

ABSTRACT. Selecting an appropriate neural architecture for a given dataset is an open problem in machine learning. Neuroevolution algorithms, such as NEAT, have shown great promise in automating this process. An extension of NEAT called EXAMM has demonstrated the capability to generate recurrent neural architectures for various time series datasets. At each iteration the evolutionary process is furthered using randomly selected mutation or crossover operations, which are chosen in accordance with pre-assigned probabilities. In this paper we present a self-adapting version of EXAMM that incorporates finite action-set learning automata (FALA), a reinforcement learning technique. FALA is used to dynamically adjust the aforementioned probabilities, thereby guiding the evolutionary process, while also significantly reducing the number of required hyperparameters. It is also demonstrated that this approach improves the performance of the generated networks with statistical significance. Furthermore, the evolution of the adapted probabilities is analyzed to gain further insight into the inner workings of EXAMM.

ABSTRACT. Neural networks are traditionally represented in terms of their weights. A key property of this representation is that there are multiple representations of a network which can be obtained by permuting the order of the neurons. These representations are generally not compatible between networks, making recombination a challenge for two arbitrary neural networks, an issue known as the "permutation problem" in neuroevolution. This paper proposes a method by which a neural network is represented in terms of interactions between neurons rather than explicit weights, and which works for both fully connected and convolutional networks. In addition to reducing the number of free parameters, this encoding is agnostic to the ordering of neurons, bypassing a key problem for direct weight-based representations. This allows us to transplant individual neurons and layers into another network without accounting for the specific ordering of neurons. We show through experiments on the MNIST and CIFAR10 datasets that this method is capable of representing networks which achieve comparable performance to direct weight representation, and that transfer done this way preserves a larger degree of performance than through direct weight transfer.

ABSTRACT. Deep reinforcement learning has met noticeable successes recently for a wide range of control problems. However, this is typically based on thousands of weights and non-linearities, making solutions complex, not easily reproducible, uninterpretable and heavy. The present paper presents genetic programming approaches. Results are competitive, in particular in the case of delayed rewards, and the solutions are lighter by orders of magnitude and much more understandable.

ABSTRACT. Model diversity is crucial for ensemble classifiers, which make predictions by combining predictions from multiple simpler models. While ensemble classifiers often outperform single-model classifiers, their success depends heavily on the ensembles construction. Genetic programming (GP) is a powerful evolutionary algorithm that can evolve populations of simple classifiers, making it seemingly an ideal technique for creating ensemble classifiers. However, in practice GP tends to produce populations of models with correlated predictions. Recent work in the broader evolutionary computing community has begun focusing on methods, such as MAP-Elites, for evolving diverse populations. In this work, we present Neuro-MAP-Elites, a novel framework for evolving error-diverse populations of classifiers.

ABSTRACT. Deep Learning Algorithms are widely implemented and have reached state-of-the-art results in several scientific investigations. In medical images domain and Computer-Assisted Detection (CAD) systems, Convolution Neural Networks (CNNs) are the preferred deep network architecture. Despite getting good results, there are still some obstacles to overcome, namely the problem of overfitting. Lately, the Data Augmentation (DA) has been integrating the training pipeline of Deep Neural Networks to mitigate those issues. The effectiveness of classical image transformations in increasing the performance of classification tasks in medical imaging domain has been reported. However, the search for a suitable augmentation strategy is performed manually. This approach, mainly made of trial-and-error tasks can be very time-consuming and complex. Thereupon, a novel data augmentation approach is proposed. The approach is an Evolutionary Machine Learning approach that is able to automatically define an optimised DA strategy for each medical image classification task. Thus, the approach combines two algorithms, an evolutionary algorithm and combined with a deep learning algorithm to find suitable DA strategies. The results obtained demonstrate that the same or better level of performance is achieved when the Transformation Functions (TFs) and their parameters are defined automatically instead of manually.

ABSTRACT. Recently it was shown, using the typical mutation mechanism that is used in evolutionary algorithms, that monotone conjunctions are provably evolvable under Bernoulli(p)^n distributions for p ∈(0, 1/3]∪{1/2}. A natural question is whether this mutation mechanism allows convergence under other distributions as well. Our experiments indicate that the answer to this question is affirmative and, at the very least, this mechanism converges under Bernoulli(p)^n distributions when p takes values outside of the known provable regime.

ABSTRACT. Grammar-Guided Genetic Programming (GGGP) has two main flavors, Context-Free Grammar GP (CFG-GP) and Grammatical Evolution (GE). GE enjoys multiple benefits, leading to being the most widely-used approach. However, GE also suffers from disadvantages.

In this paper, we first review the established advantages and disadvantages of both GE and CFG-GP. Then, we identify three new advantages of CFG-GP over GE: direct evaluation, in-node storage, and deduplication. We conclude that there is further need for studying the performance of CFG-GP and GE.

ABSTRACT. Explainability is crucial to support the adoption of machine learning as a tool for scientific discovery. In the biomedical domain, ontologies and knowledge graphs are a unique opportunity to explore domain knowledge, but most knowledge graph-based approaches employ graph embeddings, which are not explainable. However, when the prediction target is finding a relation between two entities represented in the graph, such as in the case of protein-protein interaction prediction, semantic similarity presents itself as a natural explanatory mechanism.

This work uses genetic programming over a set of semantic similarity values, each describing a semantic aspect represented in the knowledge graph, to generate global and interpretable explanations for protein-protein interaction prediction. Our experiments reveal that genetic programming algorithms coupled with semantic similarity produce global models relevant to understanding the biological phenomena.

ABSTRACT. In this paper we present an improvement over previous attempts for addressing evolutionary 4-part harmony. We describe first how the set of rules used in the fitness function, required in this kind of composition technique, has been enlarged, thus allowing augmented sixth chords, Neapolitan chord, as well as ninth chords. This greatly increases the search space. We analyse the difficulties for addressing the problem, describe some attempts to improve quality of results and finally show an example evolved using the new approach.

ABSTRACT. A ghting game is a 1 vs 1 confrontation in which the players (characters) try to reduce the health bar of the opponent by punching or kicking him. Players can normally also do special movements, which are more complicated to execute, but which can reduce the rival's health in a higher amount. Articial Intelligence (AI) engines in this type of games have a big handicap in comparison to other genres, since they must yield very fast decisions, given the highly dynamic rival movements. In this paper, we have created a Non-Player Character (or bot) aiming to beat to any opponent (being a human player or another NPC). To this end, we have started from a competitive bot (from the state of the art) having as AI engine a set of rules, depending on some conditions, threshold and weights. Then we have optimized these values by means of dierent schemes of Genetic Algorithms (GAs).

ABSTRACT. The musical instrument proposed here infers imagined music from readings of brain activity via machine learning. During training one EEG (electroencephalo- gram) cap is connected to a Variational Autoencoder (VAE) and a library of sound samples are connected to a second VAE. Time series data from the Brain and the music recorded synchronously are trained on the two separate VAEs. When the instrument is in use, a musician hears, or possibly imagines a sound, and the EEG is rendered in real-time to sample sounds from the latent music space. Mental activity is mapped directly to perceptions, delimited by the latent space of generalized sounds. The plan will dive into the Variational Autoencoder (VAE) and Regressive Flows (RF) as a mapping strategy, where RF creates an invertible map between two trained latent spaces of VAEs. In this case, one VAE contains a latent space of EEG data and the other contains sound samples.

ABSTRACT. Online Embodied Evolution is a distributed learning method for collective heterogeneous robotic swarms, in which evolution is carried out in a decentralized manner. In this work, we address the problem of promoting reproductive isolation, a feature that has been identified as crucial in situations where behavioral specialization is desired. We hypothesize that one way to allow a swarm of robots to specialize on different tasks and create situations where division of labor emerge, is through the promotion of diversity. Our contribution is twofold, we describe a method that allows a swarm of heterogeneous agents evolving online to maintain a high degree of diversity in behavioral space in which selection is based on originality. We also introduce a method to measure behavioral distances that is well suited to online distributed settings and, argue that it provides reliable measurements and comparisons. We test the claim on a concurrent foraging task and the experiments show that diversity is indeed preserved and, that different behaviors emerge in the swarm; suggesting the emergence of reproductive isolation. Finally, we employ different analysis tools from computational biology that further support this claim.

ABSTRACT. Creatures in the real world constantly encounter new and diverse challenges that they have never seen before. They will often need to adapt to some of these tasks and solve them in order to survive. This almost endless world of novel challenges is not as common in virtual environments, where artificially evolving agents often have a limited set of tasks to solve. An exception to this is the field of open-endedness where the goal is to create unbounded exploration of interesting artefacts. We want to move one step closer to creating simulated environments similar to the diverse real world, where agents can both find solvable tasks, and adapt to them. Through the use of MAP-Elites we create a structured repertoire, a map, of tasks and virtual creatures that solve them. By using novelty as a dimension in the grid, the map can continuously develop to encourage exploration of new environments. The agents must adapt to the environments found, but can also search for environments within each cell of the grid to find the one that best fits their set of skills. Our approach combines the structure of MAP-Elites, which can allow the virtual creatures to use adjacent cells as stepping stones to solve increasingly difficult environments, with open-ended innovation. This leads to a search that is unbounded, but still has a clear structure. We find that while handcrafted bounded dimensions for the map lead to quicker exploration of a large set of environments, both the bounded and unbounded approach manage to solve a diverse set of tasks.

ABSTRACT. With the recent advances of 3D printing, modular robots, and low-cost manipulators, the evolution of robots, including morphologies and controllers, has become possible to perform in a physical setup without using any simulators. In this scenario, the evolution cannot be parallelized and the wall time becomes a scare resource that should be used wisely. This paper analyses different algorithms by using the wall time as a stopping criterion for evolution, and it takes into account that wall time depends on the evaluation time plus other factors like the time to assemble and disassemble robots before and after an evaluation. Results suggest that (i) genetic algorithms are severely penalized, (ii) genetic algorithms can be improved by performing several evaluations of controllers for each morphology, and that (iii) evolutionary strategies that can chain several evaluations of robots with close morphologies can outperform other evolutionary algorithms.