ABSTRACT. We present the GENRT music generation system specifically designed for making soundtracks to fit given media such as video clips. This is based on Neo-Riemannian Theory (NRT), an analytical approach to describing chromatic chord progressions. We describe the implementation of GENRT in terms of a generative NRT formalism, which produces suitable chord sequences to fit exact timing requirements of media episodes and events. We provide details of an illustrative example producing a soundtrack to a clip from the film A Beautiful Mind.

ABSTRACT. With the introduction of musical AI in society comes the question of how it will be received by the public. An empirical study was conducted to investigate the hypotheses that human listeners hold a negative bias against computer-composed music. 163 participants were recruited from Amazon’s MTurk to fill out a survey asking participants to rank 5 computer-composed and 5 human-composed musical excerpts. Participants were split into two groups, one informed of correct authorship, the other deceived. The hypothesis, that those in the informed group would rank computer-composed excerpts as lower than human-composed excerpts, was not supported by significant results. We outline potential weaknesses in our design and present possible improvements for future work.

ABSTRACT. This paper introduces a prototype of SketchSynth, a system that enables users to graphically control synthesis using sketches of cross-modal associations between sound and shape. The development is motivated by finding alternatives to technical synthesiser controls to enable a more intuitive realisation of sound ideas. There is strong evidence that humans share cross-modal associations between sound and shapes, and recent studies found similar patterns when humans represent sound graphically. Compared to similar cross-modal mapping architectures, this prototype uses a deep classifier that predicts the character of a sound rather than a specific sound. The prediction is then mapped onto a semantically annotated FM synthesiser dataset. This approach allows for a perceptual evaluation of the mapping model and gives the possibility to be combined with various sound datasets. Two models based on architectures commonly used for sketch recognition were compared, convolutional neural networks (CNNs) and recurrent neural networks (RNNs). In an evaluation study, 62 participants created sketches from prompts and rated the predicted audio output. Both models were able to infer sound characteristics on which they were trained with over 84% accuracy. Participant ratings were significantly higher than the baseline for some prompts, but revealed a potential weak point in the mapping between classifier output and FM synthesiser. The prototype provides the basis for further development that, in the next step, aims to make SketchSynth available online to be explored outside of a study environment.

ABSTRACT. Large Language Models such as GPT-3 exhibit generative language capabilities with multiple potential applications in creative practice. In this paper, we present a method for data sonification that employs the GPT-3 model to create semantically relevant mappings be- tween artificial intelligence-generated natural language descriptions of data, and human-generated descriptions of sounds. We implemented this method in a public art installation to generate a soundscape based on data from different systems. While common sonification approaches rely on arbitrary mappings between data values and sonic values, our ap- proach explores the use of language models to achieve a mapping not via values but via meaning. We find our approach is a useful tool for musi- fication practice and demonstrates a new application of generative lan- guage models in creative new media arts practice. We show how different prompts influence data to sound mappings, and highlight that matching the embeddings of texts of different lengths produces undesired behavior.

ABSTRACT. In the classical western musical tradition, the mutual simultaneous appearance of two tones in a melody is determined by harmony, i.e. the ratio of their frequencies. To perform NLP-based methods for MIDI file analysis, one needs to construct vector embeddings of chords, taking mutual harmonicity into account. Previous works utilising this idea were based on the notion of Euler's Tonnetz. Being a beautiful topological model describing consonance relations in music, the classical Tonnetz has a certain disadvantage in that it forgets particular octaves. In this paper, we introduce the mathematical generalisation of Tonnetz taking octaves into account. Based on this model, we introduce several types of metrics on chords and use them to construct chordal embeddings. These embeddings are tested on two types of tasks: the chord estimation task, based on the Harmony Transformer model, and the music generation task, provided on the basis of TonicNet.

ABSTRACT. In this paper, we explore the prospect of accelerating tree-based genetic programming (TGP) by way of modern field-programmable gate array (FPGA) devices, which is motivated by the fact that FPGAs can sometimes leverage larger amounts of data/function parallelism, as well as better energy efficiency, when compared to general-purpose CPU/GPU systems. We compare to the recent TensorGP tool executing on a modern 8nm GPU, and we show that our accelerator implemented on a 14nm FPGA achieves an average speedup of 43x. When compared to the popular baseline tool DEAP executing across all cores of a 2-socket, 28-core (56-thread), 14nm CPU server, our accelerator achieves an average speedup of 4,902x. Finally, when compared to the recent state-of-the-art tool Operon executing on the same 2-processor CPU system, our accelerator executes about 2.4x slower on average. Despite not achieving an average speedup over every tool tested, our single-FPGA accelerator is the fastest in several instances, and we describe five future extensions that could allow for a 32-144x speedup over our current design. Overall, we estimate that a future version of our accelerator will constitute a state-of-the-art GP system for many applications.

ABSTRACT. With the recent advances in the Machine Learning field, alongside digital circuits becoming more complex each day, machine learning based methods are being used in error-tolerant applications to solve the challenges imposed by large integrated circuits, where the designer can obtain a better overall circuit while relaxing its accuracy requirement. One of these methods is the Cartesian Genetic Programming (CGP), a subclass of Evolutionary Algorithms that uses concepts from biological evolution applied in electronic design automation. CGP-based approaches show advantages in the logic learning and logic optimization processes. However, the main challenge of CGP-based flows is the extensive runtime compared to other logic synthesis strategies. We propose a new strategy to tackle this challenge, called Adaptive Batch Size (ABS) CGP, in which the CGP training incrementally exposes the search algorithm to more terms of the truth table along the evolutionary process. The proposed approach was evaluated in nine exemplars from the IWLS 2020 contest, in which 3 exemplars are from the arithmetic domain, and six are from image recognition domain, specifically three from the CIFAR-10 dataset and three from the MNIST dataset. The results show that ABS presented an accuracy increase of up to 8.19% and decreased the number of candidate solutions evaluations required by up to 84.56%, in which directly affects the runtime of the algorithm. Furthermore, for all circuits, no significant accuracy reduction was observed while a significant reduction in the number of evaluations was achieved.

ABSTRACT. In this paper, we focus on the evolutionary design of programs capable of capturing more randomness and outliers in the input data set than the common genetic programming (GP)-based methods typically allow. We propose Genetic Programming with Associative Memory (GPAM) -- a GP-based system for symbolic regression which can utilize a small associative memory to store various data points to better approximate the original data set. The method is evaluated on five standard benchmarks in which a certain number of data points is replaced by randomly generated values. In another case study, GPAM is used as an on-chip generator capable of approximating the weights for a convolutional neural network (CNN) to reduce the access to an external weight memory. Using Cartesian genetic programming (CGP), we evolved expression-memory pairs that can generate weights of a single CNN layer. If the associative memory contains 10% of the original weights, the weight generator evolved for a convolutional layer can approximate the original weights such that the CNN utilizing the generated weights shows less than 1% drop in the classification accuracy on the MNIST data set.

ABSTRACT. This paper concerns using of the code2vec vector embeddings of the source code to improve automatic source code generation in Grammatical Evolution. Focusing on a particular programming language, such as Java in the research presented, and being able to represent each Java function in the form of a continuous vector in a linear space by the Code2Vec model, GE gains some additional knowledge on similarities between constructed functions in the linear space instead of semantic similarities, which are harder to process. We propose a few improvements to the regular GE algorithm, including a code2vec-based initialization of the evolutionary algorithm and a code2vec-based crossover operator. Computational experiments confirm the efficiency of the approach proposed on a few typical benchmarks.

ABSTRACT. In many evolutionary computation systems, parent selection methods can affect, among other things, convergence to a solution. In this paper, we present a study comparing the role of two commonly used parent selection methods in evolving machine learning pipelines in an automated machine learning system called Tree-based Pipeline Optimization Tool (TPOT). Specifically, we demonstrate, using experiments on multiple datasets, that lexicase selection leads to significantly faster convergence as compared to NSGA-II in TPOT. We also compare the exploration of parts of the search space by these selection methods using a trie data structure that contains information about the pipelines explored in a particular run.

ABSTRACT. Given that real-world multi-objective optimization problems are generally constructed by combining individual functions to be optimized, it seems sensible that benchmark functions would also follow this procedure. Since the pool of functions to choose from is large and the number of function combinations increases exponentially with the number of objectives, we need a smart way to choose a reasonably sized and diverse collection of function combinations to use in benchmarking experiments. We propose a four-step approach that analyzes the landscape characteristics of all function combinations and selects only the most diverse ones to form a suite of problems. In this initial study, we test this idea on the pool of bbob functions and the case of two objectives. We provide a proof of concept for the proposed approach and its initial results. We also discuss its limitations to be addressed in future work.

ABSTRACT. Benchmarking is a key aspect of research into optimization algorithms, and as such the way in which the most popular benchmark suites are designed implicitly guides some parts of algorithm design. One of these suites is the black-box optimization benchmarking (BBOB) suite of 24 single-objective noiseless functions, which has been a standard for over a decade. Within this problem suite, different instances of a single problem can be created, which is beneficial for testing the stability and invariance of algorithms under transformations. In this paper, we investigate the BBOB instance creation protocol by considering a set of 500 instances for each BBOB problem. Using exploratory landscape analysis, we show that the distribution of landscape features across BBOB instances is highly diverse for a large set of problems. In addition, we run a set of eight algorithms across these 500 instances, and investigate for which cases statistically significant differences in performance occur. We argue that, while the transformations applied in BBOB instances do indeed seem to preserve the high-level properties of the functions, their difference in practice should not be overlooked, particularly when treating the problems as box-constrained instead of unconstrained.

ABSTRACT. The utilization of benchmarking techniques has a crucial role in the development of novel optimization algorithms, and also in performing comparisons between already existing methods. This is especially true in the field of evolutionary computation, where the theoretical performance of the method is difficult to analyze. For these benchmarking purposes, artificial (or synthetic) functions are currently the most widely used ones. In this paper, we present a collection of real-world robotics problems that can be used for benchmarking evolutionary computation methods. The proposed benchmark problems are a combination of inverse kinematics and path planning in robotics that can be parameterized. We conducted an extensive numerical investigation that encompassed solving 4{,}000 benchmark problems by seven selected metaheuristic algorithms. The results of this investigation showed that the proposed benchmark problems are quite difficult (multimodal and non-separable) and that they can be successfully used for differentiating and ranking various metaheuristics.

ABSTRACT. Dynamic algorithm selection aims to exploit the complementary of multiple optimization algorithms by switching between them during the search. While these kinds of dynamic algorithms have been shown to have potential to outperform their component algorithms, it is still unclear how this potential can best be realized. One promising approach is to make use of landscape features to enable a per-run trajectory-based switch. Here, the samples seen by the first algorithm are used to create a set of features which describe the landscape from the perspective of the algorithm. These features are then used to predict what algorithm to switch to.

In this work, we extend this per-run trajectory-based approach to consider a wide variety of potential points at which to perform the switch. We show that using a sliding window to capture the local landscape features contains information which can be used to predict whether a switch at that point would be beneficial to future performance. By analyzing the resulting models, we identify what features are most important to these predictions. Finally, by evaluating the importance of features and comparing these values between multiple algorithms, we show clear differences in the way the second algorithm interacts with the local landscape features found before the switch.

ABSTRACT. Metaheuristic navigation towards rare objective values in- stead of good objective values: is it a good idea? We will discuss the closed and open ends after presenting a successful replication study of Weise et al.’s ‘frequency fitness assignment’ for a hillClimber on the job shop scheduling problem.

ABSTRACT. The robust and reliable prediction of urban traffic provides a pathway to reducing pollution, increasing road safety and minimising infrastructure costs. The data driven modelling of vehicle flow through major cities is an inherently complex task, given the intricate topology of real life road networks, the dynamic nature of urban traffic, often disrupted by construction work and large-scale social events, and the various failures of sensing equipment, leading to discontinuous and noisy readings. It thus becomes necessary to look beyond traditional optimisation approaches and consider evolutionary methods, such as Genetic Programming (GP). We investigate the quality of GP traffic models, under both normal and anomalous conditions (such as major sporting events), at two levels: spatial, where we enhance standard GP with Transfer Learning (TL) and diversity control in order to learn traffic patterns from areas neighbouring the one where a prediction is needed, and temporal. In the latter case, we propose two implementations of GP with TL: one that employs a lag operator to skip over a configurable number of anomalous traffic readings during training and one that leverages Vectorial GP, particularly its linear algebra operators, to smooth out the effect of anomalous data samples on model prediction quality. A thorough experimental investigation conducted on central Birmingham traffic readings collected before and during the 2022 Commonwealth Games demonstrates our models’ usefulness in a variety of real-life scenarios.

ABSTRACT. Li-ion batteries play a key role in the sustainable development scenario since they allow a better management of renewable energy resources. The performance of Li-ion batteries are influenced by several factors. For this reason, accurate and reliable models of these batteries are needed, not only in the design phase, but also in operating conditions. In this paper, we present a novel approach based on Genetic Programming (GP) for the voltage prediction of a Lithium Titanate Oxide battery. The proposed approach uses a multi-objective optimization strategy. The models evolved takes in input the State-of-Charge (SoC) and provide as output the Charge rate (C-rate), which is used to evaluate the impact of the charge (or discharge) speed on the voltage. The experimental results showed that our approach is able to generate optimal candidate analytical models, where the choice of the preferred one is made by evaluating suitable metrics and imposing a sound trade-off between simplicity and accuracy. These results also proved that our GP-based behavioral modeling is more reliable and flexibile than those based on an a standard machine learning approach like a neural network.

ABSTRACT. In the present paper, we investigate the management of a fleet of electric vehicles. We propose a hybrid evolutionary approach for solving the problem of simultaneously planning the charging of electric vehicles and the assignment of electric vehicles to a set of reservations. The reservation assignment is optimized with an evolutionary algorithm while linear programming is used to compute optimal charging schedules. The evolutionary algorithm uses an indirect encoding and a problem-specific crossover operator. Furthermore, we propose the use of a surrogate fitness function. Experimental results on problem instances with up to 100 vehicles and 1600 reservations show that the proposed approach is able to notably outperform two approaches based on mixed integer linear programming.

ABSTRACT. In this paper, we address a problem in the field of hydraulics which is also relevant in terms of sustainability. Hydraulic jump is a physical phenomenon that occurs both for natural and man-made reasons. Its importance relies on the exploitation of the intrinsic energy dissipation characteristics and on the other hand the danger that might produce on bridges and river structures as a consequence of the interaction with the large vortex structures that are generated. In the present work, we try to address the problem of estimating the hydraulic jump roller length, whose evaluation is inherently affected by empirical errors related to its dissipative nature. This problem is based on a regression problem by exploiting a dataset of observations. Regression is performed by minimising the loss function using ten different black-box optimisers. In particular, we selected some of the most used metaheuristics, such as Evolution Strategies, Particle Swarm Optimisation, Differential Evolution and others. Furthermore, an experimental analysis has been conducted to validate the proposed approach and compare the effectiveness of the ten metaheuristics.

ABSTRACT. Decarbonization of economy is drive a world level as a principal action for reduction greenhouse effect gases emission and mitigate climatic change. One of the way decarbonization of economy is electrification of economic sectors and in this case, implementation of microgrid in different economic sectors as households, industry and commerce is a great mechanism that allow integrate renewable energy into the power system and to contribute with accelerated energy transition for decarbonization. However, microgrid include self-generation by renewable energy and distributed generation too, as well as energy efficiency in the consumer. Microgrid have several benefits, such as, energetic, economic, and environmental benefits for user and power system but to security energy supply it is necessary to balance the offer and demand of electricity in all moments, which in this case needs to be estimated for market of the next day. The problem here, it is as estimated generation and consume for next day? when determinant of offer and demand are variable. For this, the paper proposes algorithms of forecasting based in machine learning whit good results in a decision support system of management of energy for microgrid and its integration with conventional power system distribution.

ABSTRACT. Fitness Landscape Analysis (FLA) for loss/gain functions for Machine Learning is an emerging research trend in Computational Intelligence that offered an alternative view on how learning algorithms work and should be designed. The vast majority of these recent studies investigate supervised learning whereas reinforcement learning remains so far nearly unaddressed. This paper performs a FLA on the reinforcement learning of a deep neural network for a simulated robot control task and focuses on ruggedness and neutrality. Two configurations of the physical system under investigation are considered and studied separately to highlight differences and similarities. Furthermore, the results of the performed FLA are put into relation with performance of the learning algorithm with the aim of achieving an understanding of most suitable parameter setting. Numerical results indicate a correlation between ruggedness and exploration to enable more optimal reinforcement learning. In the presence of high ruggedness the algorithm displays its best performance when the control parameters are set to enable a high degree of exploration. Conversely, when the landscape appears less rugged a less exploratory behaviour seems to contribute to the best performance of the learning algorithm.

ABSTRACT. As a scalable and lightweight infrastructure technology, containers are quickly gaining popularity in cloud data centers. However, dynamic Resource-Allocation in Container-based clouds (RAC) is challenging due to two interdependent allocation sub-problems, allocating dynamic arriving containers to appropriate Virtual Machines (VMs) and allocating VMs to multiple Physical Machines (PMs). Most of existing research works assume homogeneous PMs and rely on simple and manually designed heuristics such as Best Fit and First Fit, which can only capture limited information, affecting their effectiveness of reducing energy consumption in data centers. In this work, we propose a novel hybrid Cooperative Coevolution Genetic Programming (CCGP) hyper-heuristic approach to automatically generate heuristics that are effective in solving the dynamic RAC problem. Different from existing works, our approach hybridizes Best Fit to automatically designed heuristics to coherently solve the two interdependent sub-problems. Moreover, we introduce a new energy model that accurately captures the energy consumption in a more realistic setting than that in the literature, e.g., real-world workload patterns and heterogeneous PMs. The experiment results show that our approach can significantly reduce energy consumption, in comparison to two state-of-the-art methods.

ABSTRACT. Deep-neuroevolution is the optimisation of deep neural architecture using evolutionary computation. Amongst these techniques, Fast-Deep Evolutionary Network Structured Representation (Fast-DENSER) has achieved considerable success in the development of Convolutional Neural Networks (CNNs) for image classification. In this study, variants of this algorithm are seen through the lens of Neuroevolution Trajectory Networks (NTNs), which uses complex networks modelling and visualisation to uncover intrinsic characteristics. We examine how evolution uses previously acquired knowledge on some datasets to inform the search for new domains with a specific focus on the architecture of CNNs. Results show that the transfer learning paradigm works as intended as networks mutate, incorporating layers from the best models trained on previous datasets. The use of specifically designed NTNs in this analysis enabled us to perceive the architectural characteristics that evolution favours in the design of CNNs. These findings provide novel insights that may inform the future creation of Deep Neural Networks (DNNs).

ABSTRACT. Operational risk is the risk associated with business operations in an organisation. With respect to agricultural crops, in particular, operational risk is a fundamental concept for establishing a differentiated coverage and when seeking protection against different risks. For cultivation, these risks are related to the agricultural business process and also to external risk events. An operational risk assessment allows for identifying limits of environmental and financial sustainability. Specifically, in oil palm cultivation, the characterisation of the associated risk remains a challenge from a technological perspective. In order to advance in this direction, researchers have used different technologies, including spectral aerial images, unmanned aerial vehicles to construct a vegetation index, augmented intelligent platforms for real-time monitoring, and adaptive fuzzy models to estimate the operational risk. In line with these technological developments, in this paper we propose a framework for the estimation of the risk assessment associated with the disease of Lethal Wilt (LW) in oil palm plantations. Although our purpose is not to predict lethal wilt, since the framework starts from the result of a prediction model, a model to detect LW in an early stage is used for the demonstration. For the implementation of the prediction model, we use a novel deep learning system, based on two neural networks. This refers to a case study conducted at UNIPALMAS. We show that the appropriateness of our system aims to evaluate LW operational risks with a confidence level of 99.9\% and for a period of 6 months during.

ABSTRACT. This paper describes a new memetic semantic algorithm for symbolic regression (SR). While memetic-computation offers a way to encode domain-knowledge into a population-based process, whereas semantic-based algorithms allow one to further improve them locally to achieve a desired output. Hence, combining memetic and semantic enables us to enhance the exploration and exploitation feature of genetic programming (GP), as well as, to discover short symbolic expressions that are easy to understand and interpret without the expressivity characteristics of symbolic regression. Experimental results showed that our proposed memetic semantic algorithm can outperform traditional evolutionary and non-evolutionary methods on several real-world datasets, paving a new direction to handle both the bloating and generalization endeavors of genetic programming.

ABSTRACT. Denoising Autoencoder Genetic Programming (DAE-GP) is a model-based evolutionary algorithm that uses denoising autoencoder long short-term memory networks as probabilistic model to replace the standard recombination and mutation operators of genetic programming (GP).In this paper, we use the DAE-GP to solve a set of nine standard real-world symbolic regression tasks. We compare the prediction quality of the DAE-GP to standard GP, geometric semantic GP (GSGP), and the gene-pool optimal mixing evolutionary algorithm for GP (GOMEA-GP), and find that the DAE-GP shows similar prediction quality using a much lower number of fitness evaluations than GSGP or GOMEA-GP. In addition, the DAE-GP consistently finds small solutions. The best candidate solutions of the DAE-GP are 69\% smaller (median number of nodes) than the best candidate solutions found by standard GP.An analysis of the bias of the selection and variation step for both the DAE-GP and standard GP gives insight into why differences in solution size exist: the strong increase in solution size for standard GP is a result of both selection and variation bias.The results highlight that learning and sampling from a probabilistic model is a promising alternative to classic GP variation operators where the DAE-GP is able to generate small solutions for real-world symbolic regression tasks.

ABSTRACT. We consider how generative AI systems could and should evolve from being used as co-creative tools for artefact generation to co-creative collaborators for enhancing cultural knowledge, via artefact generation. We argue that while generative deep learning techniques and ethos have many drawbacks in this respect, neuro-symbolic approaches and increased AI autonomy could improve matters and these techniques so that AI systems may be able to add to cultural knowledge directly. To do this, we consider what cultural knowledge is, differences between scientific and artistic applications of generative AI, stakeholders and ethical issues. We also present a case study in decision foregrounding for generative music.

ABSTRACT. In this work, we create artistic closed loop curves that trace out images and 3D shapes, which we then hide in musical audio as a form of steganography. We use traveling salesperson art to create artistic plane loops to trace out image contours, and we use Hamiltonian cycles on triangle meshes to create artistic space loops that fill out 3D surfaces. Our embedding scheme is designed to faithfully preserve the geometry of these loops after lossy compression, while keeping their presence undetectable to the audio listener. To accomplish this, we hide each dimension of the curve in a different frequency, and we perturb a regularized version of the magnitude of that frequency to best match the target curve at that dimension, while hiding scale information in that frequency's phase. In the process, we exploit geometric properties of the curves to help to more effectively hide and recover them. Our scheme is simple and encoding happens efficiently with a nonnegative least squares framework, while decoding is trivial. We validate our technique quantitatively on large datasets of images and audio, and we show results of a crowd sourced listening test that validate that the hidden information is indeed unobtrusive.

ABSTRACT. Electronic music artists and sound designers have unique workflow practices that necessitate specialized approaches for developing music information retrieval and creativity support tools. Furthermore, electronic music instruments, such as modular synthesizers, have near-infinite possibilities for sound creation and can be combined to create unique and complex audio paths. The process of discovering interesting sounds is often serendipitous and impossible to replicate. For this reason, many musicians in electronic genres record audio output at all times while they work in the studio. Subsequently, it is difficult for artists to rediscover audio segments that might be suitable for use in their compositions from thousands of hours of recordings. In this paper, we describe a creative tool for musicians to rediscover their previous recordings, re-contextualize them with other recordings, and create original live music compositions in real-time. A bi-modal AI-driven approach uses generated lyric lines to find compatible audio clips from the artist's past studio recordings, and uses them to generate new lyric lines, which in turn are used to find other clips, thus creating a continuous and evolving stream of music and lyrics. The intent is to keep the artists in a state of creative flow conducive to music creation rather than taking them into an analytical/critical state of deliberately searching for past audio segments. The system can run in either a fully autonomous mode without user input, or in a live performance mode, where the artist plays live music, while the system "listens" and creates a continuous stream of music and lyrics in response.

ABSTRACT. This paper introduces SUNMASK, an approach for generative sequence modeling based on masked unrolled denoising autoencoders. By explicitly incorporating a conditional masking variable, as well as using this mask information to modulate losses during training based on expected exemplar difficulty, SUNMASK models discrete sequences without direct ordering assumptions. The addition of masking terms allows for fine-grained control during generation, starting from random tokens and a mask over subset variables, then predicting tokens which are again combined with a subset mask for subsequent repetitions. This iterative process gradually improves token sequences toward a structured output, while guided by proposal masks. The broad framework for unrolled denoising autoencoders is largely independent of model type, and we utilize both transformer and convolution based architectures in this work. We demonstrate the efficacy of this approach both qualitatively and quantitatively, applying SUNMASK to generative modeling of symbolic polyphonic music, and language modeling for English text.

ABSTRACT. This paper deals with the problem of deinterleaving a sequence of signals received from different emitters at different time steps. It is assumed that this pulse sequence can be modeled by a collection of processes over disjoint finite sub-alphabets, which have been randomly interleaved by a switch process. A known method to solve this problem is to maximise the likelihood of the model which involves a partitioning problem of the whole alphabet. This work presents a new memetic algorithm using a dedicated likelihood-based crossover to efficiently explore the space of possible partitions. The algorithm is first evaluated on synthetic data generated with Markov processes, then its performance is assessed on electronic warfare datasets.

ABSTRACT. This work presents a hyper-heuristic approach with online learning, which combines Monte Carlo Tree Search with multiple local search operators selected on the fly during the search. The impacts of different operator policies, including proportional bias, one-armed bandit, and deep learning methods, are investigated. Experiments on well-known benchmarks of the Weighted Vertex Coloring Problem are conducted to highlight the advantages and limitations of each dynamic selection strategy.

ABSTRACT. The design of binary error-correcting codes is a challenging optimization problem with several applications in telecommunications and storage, which has also been addressed with metaheuristic techniques and evolutionary algorithms. Still, all these efforts focused on optimizing the minimum distance of unrestricted binary codes, i.e., with no constraints on their linearity, which is a desirable property for efficient implementations. In this paper, we present an Evolutionary Strategy (ES) algorithm that explores only the subset of linear codes of a fixed length and dimension. To that end, we represent the candidate solutions as binary matrices and devise variation operators that preserve their ranks. Our experiments show that up to length $n=14$, our ES always converges to an optimal solution with a full success rate, and the evolved codes are all inequivalent to the Best-Known Linear Code (BKLC) given by MAGMA. On the other hand, for larger lengths, both the success rate of the ES as well as the diversity of the evolved codes start to drop, with the extreme case of $(16,8,5)$ codes which all turn out to be equivalent to MAGMA's BKLC.

ABSTRACT. Evolutionary algorithms make countless random decisions during selection, mutation and crossover operations. These random decisions require a steady stream of random numbers. We analyze the expected number of random bits used throughout a run of an evolutionary algorithm and refer to this as the cost of randomness. We give general bounds on the cost of randomness for mutation-based evolutionary algorithms using 1-bit flips or standard mutations using either a naive or a common, more efficient implementation that uses $\Theta(\log n)$ random bits per mutation.

Uniform crossover is a potentially wasteful operator as the number of random bits used equals the Hamming distance of the two parents, which can be up to~$n$. However, we show for a (2+1) Genetic Algorithm that is known to optimize the test function OneMax in roughly $(e/2)n \ln n$ expected evaluations, twice as fast as the fastest mutation-based evolutionary algorithms, that the total cost of randomness during all crossover operations on OneMax is only $\Theta(n)$. Consequently, the use of crossover can reduce the cost of randomness below that of the fastest evolutionary algorithms that only use standard mutations.

ABSTRACT. We propose the application of a recently introduced version of ant colony optimization---negative learning ant colony optimization---to the far from most string problem. This problem is a notoriously difficult combinatorial optimization problem from the group of string selection problems. The proposed algorithm makes use of negative learning in addition to the standard positive learning mechanism in order to achieve better guidance for the exploration of the search space. In addition, we compare different versions of our algorithm characterized by the use of different objective functions. The obtained results show that our algorithm is especially successful for instances with specific characteristics. Moreover, it becomes clear that none of the existing state-of-the-art methods is best for all problem instances.

ABSTRACT. Beam Search (BS) is a popular incomplete breadth-first search widely used to find near-optimal solutions to hard combinatorial optimization problems in limited time. Its central component is an evaluation function that estimates the quality of nodes encountered on each level of the search tree. While this function is usually manually crafted for a problem at hand, we propose a Policy-Based Learning Beam Search (P-LBS) that learns a policy to select the most promising nodes at each level offline on representative random problem instances in a reinforcement learning manner. In contrast to an earlier learning beam search, the policy function is realized by a neural network (NN) that is applied to all the expanded nodes at a current level together and does not rely on the prediction of actual node values. Different loss functions suggested for beam-aware training in an earlier work, but there only theoretically analyzed, are considered and evaluated in practice on the well-studied Longest Common Subsequence (LCS) Problem. To keep P-LBS scalable to larger problem instances, a bootstrapping approach is further proposed for training. Results on established sets of LCS benchmark instances show that P-LBS with loss functions ``upper bound'' and ``cost-sensitive margin beam'' is able to learn suitable policies for BS such that results highly competitive to the state-of-the-art can be obtained.

ABSTRACT. We study the (1:s+1) success rule for controlling the population size of the (1,Lambda)-EA. It was shown by Hevia Fajardo and Sudholt that this parameter control mechanism can run into problems for large s if the fitness landscape is too easy. They conjectured that this problem is worst for the ONEMAX benchmark, since in some well-established sense ONEMAX is known to be the easiest fitness landscape. In this paper we disprove this conjecture and show that ONEMAX is not the easiest fitness landscape with respect to finding improving steps. As a consequence, we show that there exist s and epsilon such that the self-adjusting (1,Lambda)-EA with (1:s+1)-rule optimizes ONEMAX efficiently when started with epsilon*n zero-bits, but does not find the optimum in polynomial time on DYNAMIC BINVAL. Hence, we show that there are landscapes where the problem of the (1:s+1)-rule for controlling the population size of the (1,Lambda)-EA is more severe than for ONEMAX.

ABSTRACT. Autoencoders are powerful models for non-linear dimensionality reduction. However, their neural network structure makes it difficult to interpret how the high dimensional features relate to the low-dimensional embedding, which is an issue in applications where explainability is important. There have been attempts to replace both the neural network components in autoencoders with interpretable genetic programming (GP) models. However, for the purposes of interpretable dimensionality reduction, we observe that replacing only the encoder with GP is sufficient. In this work, we propose the Genetic Programming Encoder for Autoencoding (GPE-AE). GPE-AE uses a multi-tree GP individual as an encoder, while retaining the neural network decoder. We demonstrate that GPE-AE is a competitive non-linear dimensionality reduction technique compared to conventional autoencoders and a GP based method that does not use an autoencoder structure. As visualisation is a common goal for dimensionality reduction, we also evaluate the quality of visualisations produced by our method, and highlight the value of functional mappings by demonstrating insights that can be gained from interpreting the GP encoders.

ABSTRACT. Evolutionary ensemble learning methods with Genetic Programming have achieved remarkable results on regression and classification tasks by employing quality-diversity optimization techniques like MAP-Elites and Neuro-MAP-Elites. The MAP-Elites algorithm uses dimensionality reduction methods, such as variational auto-encoders, to reduce the high-dimensional semantic space of genetic programming to a two-dimensional behavioral space. Then, it constructs a grid of high-quality and diverse models to form an ensemble model. In MAP-Elites, however, variational auto-encoders rely on Euclidean space topology, which is not effective at preserving high-quality individuals. To solve this problem, this paper proposes a principal component analysis method based on a cosine-kernel for dimensionality reduction. In order to deal with unbalanced distributions of good individuals, we propose a zero-cost reference points synthesizing method. Experimental results on 108 datasets show that combining principal component analysis using a cosine kernel with reference points significantly improves the performance of MAP-Elites evolutionary ensemble learning algorithm.

ABSTRACT. An approach to evolutionary ensemble learning for classification is proposed in which boosting is used to construct a stack of programs. Each application of boosting identifies a single champion and a residual dataset, i.e. the training records that thus far were not correctly classified. The next program is only trained against the residual, with the process iterating until some maximum ensemble size or no further residual remains. Training against a residual dataset actively reduces the cost of training. Deploying the ensemble as a stack also means that only one classifier might be necessary to make a prediction, so improving interpretability. Benchmarking studies are conducted to illustrate competitiveness with the prediction accuracy of current state-of-the-art evolutionary ensemble learning algorithms, while providing solutions that are orders of magnitude simpler. Further benchmarking with a high cardinality dataset indicates that the proposed method is also more accurate and efficient than XGBoost.

ABSTRACT. The field of video games is of great interest to researchers in computational intelligence due to the complex, rich and dynamic nature they provide. Evolutionary algorithms have been used in this field with satisfactory results.

In this paper we propose the use of Genetic Programming to generate an agent that plays the \textit{Bomberman} game. This game is very interesting to deal with due to its challenging and asynchronous attack nature. Because we have not encountered agents already created during the development of this work the use of co-evolution is proposed to generate versatile agents without the need to overfit with previously created enemies. The use of lexicographic fitness is also proposed to force the creation of more aggressive agents. The results show that this method can generate agents good enough to face in artificial intelligence competitions.

ABSTRACT. Collective decision-making is a process that allows a group of autonomous agents to make a decision in a way that once the decision is made it cannot be attributed to any agent in the group. In the swarm robotics literature, collective decision-making mechanisms have generally been designed using behaviour-based control structures. That is, the individual decision-making mechanisms are integrated into modular control systems, in which each module concerns a specific behavioural response required by the robots to respond to physical and social stimuli. Recently, an alternative solution has been proposed which is based on the use of dynamical neural networks as individual decision-making mechanisms. This alternative solution proved effective in a perceptual discrimination task under various operating conditions and for swarms that differ in size. In this paper, we further investigate the characteristics of this neural model for opinion selection using three different tests. The first test examines the ability of the neural model to underpin consensus among the swarm members in an environment where all available options have the same quality and cost (i.e., a symmetrical environment). The second test evaluates the neural model with respect to a type of environmental variability related to the spatial distribution of the options. The third test examines the extent to which the neural model is tolerant to the failure of individual components. The results of our simulations show that the neural model allows the swarm to reach consensus in a symmetrical environment, and that it makes the swarm relatively resilient to major sensor failure. We also show that the swarm performance drops in accuracy in those cases in which the perceptual cues are patchily distributed.

ABSTRACT. Designing controllers for a swarm of robots such that collaborative behaviour emerges at the swarm level is known to be challenging. Evolutionary approaches have proved promising, with attention turning more recently to evolving repertoires of diverse behaviours that can be be used to compose heterogeneous swarms or mitigate against faults. Here we extend existing work by combining a Quality-Diversity algorithm (MAP-Elites) with a genetic-programming (GP) algorithm to evolve repertoires of behaviour-trees that define the robot controllers. We compare this approach two variants of GP, one of which uses an implicit diversity method. Our results show that QD approach results in larger and more diverse repertoires than the other methods with no loss in quality with respect to the best solutions found. Given that behaviour-trees have the added advantage of being human-readable compared to neural controllers that are typically evolved, the results provide a solid platform for future work in composing heterogeneous swarms.

ABSTRACT. The traffic light scheduling problem is undoubtedly one of the most critical problems in a modern traffic management system. Appropriate traffic light planning can improve traffic flows, reduce vehicles' emissions, and provide benefits for the whole city. Metaheuristics, notably the Cellular Genetic Algorithm (cGA), offer an alternative way of solving this optimization problem by providing "good solutions" to adjust the traffic lights to mitigate traffic congestion. However, one of the unresolved issues is these methods use very time-consuming operations. Specifically, the evaluation is a complex process since a simulator should be executed to get the quality of the solutions. In this work, we focus on this topic and propose using an artificial neural network (as a surrogate system) to tackle this problem. Our experiments show very promising results since our proposal can significantly reduce the execution time while maintaining (and even, in some scenarios, improving) the quality of the solutions.

ABSTRACT. To date, several algorithms have been proposed to deal with constrained optimization problems, particularly multi-objective optimization problems (MOOPs), in real-world engineering. This work extends the 2020 study by Gandomi and Deb on boundary updating (BU) for the MOOPs. The proposed method is an implicit constraint handling technique (CHT) that aims to cut the infeasible search space, so the optimization algorithm focuses on feasible regions. Furthermore, the proposed method is coupled with an explicit CHT, namely, feasibility rules and then the search operator (here NSGA-II) is applied to the optimization problem. To illustrate the applicability of the proposed approach for MOOPs, a numerical example is presented in detail. Additionally, an evaluation of the BU method was conducted by comparing its performance to an approach without the BU method while the feasibility rules (as an explicit CHT) work alone. The results show that the proposed method can significantly boost the solutions of constrained multi-objective optimization.

ABSTRACT. We consider the issue of intensification/diversification balance in the context of a memetic algorithm for the multiobjective optimization of investment portfolios with cardinality constraints. We approach this issue in this work by considering the selective application of knowledge-augmented operators (local search and a memory of elite solutions) based on the search epoch in which the algorithm finds itself, hence alternating between unbiased search (guided uniquely by the built-in search mechanics of the algorithm) and focused search (intensified by the use of the problem-aware operators). These operators exploit Sharpe index (a measure of the relationship between return and risk) as a source of problem knowledge. We have conducted a sensibility analysis to determine in which phases of the search the application of these operators leads to better results. Our findings indicate that the resulting algorithm is quite robust in terms of parameterization from the point of view of this problem-specific indicator. Furthermore, it is shown that not only can other non-memetic counterparts be outperformed, but that there is a range of parameters in which the MA is also competitive when not better in terms of standard multiobjective performance indicators.

ABSTRACT. Personalised medicine (PM) is a new area of healthcare that has shown promising results in offering treatments for rare and advanced stage diseases. Nonetheless, their supply chain differs from the traditional healthcare model by adding a high level of complexity through product individualisation. The patient is now also the donor, and becomes part of a complex manufacturing and delivery system. PM biopharmaceuticals are cryopreserved (frozen) and re-engineered at specialised facilities, before being returned to the same patient. The corresponding facility location problem (FLP) consists of both manufacturing and cryopreservation sites, having a set of constraints that are not met in other healthcare supply chains. As a result, the FLP in the context of PM has only been partially analysed and additional research is still necessary to reach optimal network configurations. In this paper, we extend the solution methods previously proposed for PM FLP by using a stage-wise approach in which the (constrained binary multi-objective) problem is divided into smaller, logical parts. We approach the problem from the perspective of the decision maker (DM) and use the R-NSGA-II algorithm to find a set of desirable solutions. By optimising only part of the decision space at a time, we reduce the complexity of the problem and allow the DM to compare the objective values obtained between different supply chain configurations. Our results suggest that allowing the DM to interact with the optimisation process can lead to good and desirable solutions in a shorter computational time, and more flexible network configurations.

ABSTRACT. Machine learning models can use information from gene expressions in patients to efficiently predict the severity of symptoms for several diseases. Medical experts, however, still need to understand the reasoning behind the predictions before trusting them. In their day-to-day practice, physicians prefer using \textit{gene expression profiles}, consisting of a discretized subset of all data from gene expressions: in these profiles, genes are typically reported as either over-expressed or under-expressed, using discretization thresholds computed on data from a healthy control group. A discretized profile allows medical experts to quickly categorize patients at a glance. Building on previous works related to the automatic discretization of patient profiles, we present a novel approach that frames the problem as a multi-objective optimization task: on the one hand, after discretization, the medical expert would prefer to have as few different profiles as possible, to be able to classify patients in an intuitive way; on the other hand, the loss of information has to be minimized. Loss of information can be estimated using the performance of a classifier trained on the discretized gene expression levels. We apply a state-of-the-art evolutionary multi-objective algorithm, NSGA-II, to the discretization of a dataset of COVID-19 patients that developed either mild or severe symptoms. The results show not only that the solutions found by the approach dominate traditional discretization based on statistical analysis and are more generally valid than those obtained through single-objective optimization, but that the candidate Pareto-optimal solutions preserve the sense-making that practitioners find necessary to trust the results.