EVOSTAR2024: EVOSTAR
PROGRAM FOR WEDNESDAY, APRIL 3RD
Days:
next day
all days

View: session overviewtalk overview

11:00-13:35 Session 3A: EvoMUSART 1 - Computational Intelligence for Audio-visual Creativity
Location: Room A
11:00
Enough is Enough: Learning to Stop in Generative Systems

ABSTRACT. Gene regulatory networks (GRNs) have been used artificially to drive generative systems. These systems must begin and then stop generation or growth akin to their biological counterpart. In nature, this process is controlled automatically as an organism reaches its mature form; in evolved generative systems, this is more typically controlled by hard-coded limits, which can be difficult to determine. Removing parameters from the evolutionary process and allowing stopping to occur naturally within an evolved system would allow for more natural and regulated growth. This paper illustrates that, within the appropriate context, the introduction of memory components into GRNs allows a stopping criterion to emerge. A Long Short-Term Memory style network was implemented as a GRN for an Evo-Devo generative system and was tested on one simple (single point target), and two more complex (point clouds) problems with and without structure. The novel LSTM-GRN performed well in simple tasks to optimise stopping conditions but struggled to manage more complex environments. This early work in self-regulating growth will allow for further research in more complex systems to allow the removal of hyperparameters and allowing the evolutionary system to stop dynamically and prevent organisms overshooting the optimal.

11:25
Co-creative orchestration of Angeles with layer scores and orchestration plans

ABSTRACT. Orchestration is the process of creating music for a group of instruments, combining, blending, and contrasting their sounds to produce a unique orchestral texture. In this research/creation project, our team was commissioned an AI/human orchestration of two movements of an existing piano composition. The project turned out as a perfect case study for computational creativity in music and orchestration, where the role of the model is between AI as a colleague and AI as a tool. By modeling a layer score and an orchestration plan in the orchestration process, we implement a simple Markov model that selects possible instrumentations for each score segment. Personalization of the AI and AI/human interaction occur through human segmentation of the score at two stages of the process (layer score, orchestral segments with loudness profile), through instrumentation presets, and finally through selection of the final orchestral plan and through the actual orchestration. We detail the research aspects of this co-creative project and analyze the roles of the actors involved in the creation of the final piece: the Music Information Retrieval (MIR) researchers, the orchestrator, and the algorithms.

11:50
On the impact of directed mutation applied to Evolutionary 4-part harmony models

ABSTRACT. This paper analyzes the difficulty of finding solutions for the 4-part harmony problem, both from the point of view of the size of the search space and the time required to run the algorithm. These considerations led to improved running time through parallelization, precalculation of the fitness function, and directed mutation, which reduces the time to solution. Moreover, we show how combining these techniques allows us to extend the improvements to different harmony models, including new synthetic ones that may be proposed.

12:15
Towards Sound Innovation Engines Using Pattern-Producing Networks and Audio Graphs

ABSTRACT. This study draws on the challenges that composers and sound designers face in creating and refining new tools to achieve their musical goals. Utilising evolutionary processes to promote diversity and foster serendipitous discoveries, we propose to automate the search through uncharted sonic spaces for sound discovery. We argue that such diversity promoting algorithms can bridge a technological gap between the theoretical realisation and practical accessibility of sounds. Specifically, in this paper we describe a system for generative sound synthesis using a combination of Quality Diversity (QD) algorithms and a discriminative model, inspired by the Innovation Engine algorithm. The study explores different configurations of the generative system and investigates the interplay between the chosen sound synthesis approach and the discriminative model. The results indicate that a combination of Compositional Pattern Producing Network (CPPN) + Digital Signal Processing (DSP) graphs coupled with Multi-dimensional Archive of Phenotypic Elites (MAP-Elites) and a deep learning classifier can generate a substantial variety of synthetic sounds. The study concludes by presenting the generated sound objects through an online explorer and as rendered sound files. Furthermore, in the context of music composition, we present an experimental application that showcases the creative potential of our discovered sounds.

12:40
Generating Smooth Mood-Dynamic Playlists with Audio Features and KNN

ABSTRACT. Users curate music playlists for many purposes, including focus, enjoyment and therapy. Popular music streaming services generate playlists automatically which are constant in genre or mood. We propose a method to automatically create playlists dynamic in both the Arousal-Valence emotion space and the audio features of songs. Our playlist algorithm uses a two-stage approach to sequentially choose songs, employing a K-Nearest Neighbors (KNN) model to gather potential songs based on emotion and analyzing them with acoustic similarity metrics. We developed a testing protocol which generates playlists that traverse both Arousal-Valence and audio feature spaces to evaluate the effectiveness of various audio feature data, KNN parameters, and similarity metrics. We define evaluation metrics to measure a playlist's smoothness and evenness using the Pearson correlation coefficient between dimensions and the variance of steps between songs, respectively. Our algorithm successfully creates smooth and evenly-spaced playlists that transition cohesively in both mood and genre. We explore how the choice of audio feature data, similarity metric, and KNN parameters all have an effect on playlists' smoothness and evenness across these two spaces.

13:05
Pruning Worlds into Stories: Affective Interactions as Fitness Function

ABSTRACT. An important challenge when trying to find a story to tell about some set of events that has already happened is to identify the elements in that set of events that will make a story that moves the intended audience. One possible criterion is to consider events that involve significant changes in the emotional relations between the characters involved. The present paper explores a computational model of this particular approach to the task of storytelling. An evolutionary solution is used to explore the logs of an agent-based social simulation, using metrics on the evolution of affinity between characters as fitness function, to identify sequences of events that might be good candidates for moving stories

11:00-13:35 Session 3B: EML 1 - Computational Models and Short Talks
Location: Room B
11:00
Towards Physical Plausibility in Neuroevolution Systems

ABSTRACT. The increasing usage of Artificial Intelligence (AI) models, especially Deep Neural Network (DNN), is increasing the power consumption during training and inference, posing environmental concerns and driving the need for more energy-efficient algorithms and hardware solutions. This work addresses the growing energy consumption problem in Machine Learning (ML), particularly during the inference phase. Even a slight reduction in power usage can lead to significant energy savings, benefiting users, companies, and the environment. Our approach focuses on maximizing the accuracy of Artificial Neural Network (ANN) models using a neuroevolutionary framework whilst minimizing their power consumption. To do so, power consumption is considered in the fitness function. We introduce a new mutation strategy that stochastically reintroduces modules of layers, with power-efficient modules having a higher chance of being chosen. We introduce a novel technique that allows training two separate models in a single training step whilst promoting one of them to be more power efficient than the other while maintaining similar accuracy. The results demonstrate a reduction in power consumption of ANN models by up to 29.2% without a significant decrease in predictive performance.

11:25
Evolving Reservoirs for Meta Reinforcement Learning

ABSTRACT. Animals often demonstrate a remarkable ability to adapt to their environments during their lifetime. They do so partly due to the evolution of morphological and neural structures. These structures capture features of environments shared between generations to bias and speed up lifetime learning. In this work, we propose a computational model for studying a mechanism that can enable such a process. We adopt a computational framework based on meta reinforcement learning as a model of the interplay between evolution and development. At the evolutionary scale, we evolve reservoirs, a family of recurrent neural networks that differ from conventional networks in that one optimizes not the synaptic weights, but hyperparameters controlling macro-level properties of the resulting network architecture.  At the developmental scale, we employ these evolved reservoirs to facilitate the learning of a behavioral policy through Reinforcement Learning (RL). Within an RL agent, a reservoir encodes the environment state before providing it to an action policy.We evaluate our approach on several 2D and 3D simulated environments. Our results show that the evolution of  reservoirs can improve the learning of diverse challenging tasks. We study in particular three hypotheses: the use of an architecture combining reservoirs and reinforcement learning could enable (1) solving tasks with partial observability, (2) generating oscillatory dynamics that facilitate the learning of locomotion tasks, and (3) facilitating the generalization of learned behaviors to new tasks unknown during the evolution phase.

11:50
Leveraging More of Biology in Evolutionary Reinforcement Learning
PRESENTER: Bruno Gašperov

ABSTRACT. In this paper, we survey the use of additional biologically inspired mechanisms, principles, and concepts in the area of evolutionary reinforcement learning (ERL). While recent years have seen the emergence of a swath of metaphor-laden approaches, they mostly represent reiterations of old algorithms under novel metaphors. At the same time, many promising ideas from evolutionary biology and related areas, ripe for exploitation in evolutionary machine learning, remain in relative obscurity. To rectify this, we provide a comprehensive analysis of innovative, often unorthodox approaches in ERL that utilize additional bio-inspired elements. Moreover, we identify research directions in the field with the largest potential to yield fruitful results and discuss classes of problems that could benefit the most from such research.

12:15
An Evolutionary Deep Learning Approach for Efficient Quantum Algorithms Transpilation

ABSTRACT. Gate-based quantum computation describes algorithms as quantum circuits. These can be seen as a set of quantum gates acting on a set of qubits. To be executable, the circuit requires complex transformations to comply with the physical constraints of the machines. This process is known as transpilation, where qubits' initialisation is one of its first and most challenging steps, usually done by considering the device error properties. As the size of the quantum algorithm increases, the transpilation becomes increasingly complex and time-consuming. This constitutes a bottleneck towards agile, fast, and error-robust quantum computation. This work proposes an evolutionary deep neural network that learns the qubits' initialisation of the most advanced and complex IBM heuristic used in today's quantum machines. The aim is to progressively replace weakly scalable transpilation heuristics with machine learning models. Previous work using machine learning models for qubits' initialisation suffers from some shortcomings in the proposal's correctness and generalisation as well as benchmarks diversity, utility, and availability. The present work solves those flaws by (I) devising a complete Machine Learning pipeline including the ETL component and the evolutionary deep neural model using the linkage learning algorithm P3, (II) a modeling applicable to any quantum algorithm with a special interest to both optimisation and machine learning ones, (III) diverse and fresh benchmarks using calibration data of four real IBM quantum computers collected over 10 months (Dec. 2022 and Oct. 2023) and training dataset built using four types of quantum optimisation and machine learning algorithms, as well as random ones. The proposal has been proven to be more efficient and simple than state-of-the-art deep neural models in the literature.

12:25
A Comparative Analysis of Evolutionary Adversarial One-Pixel Attacks

ABSTRACT. Adversarial attacks pose significant challenges to the robustness of machine learning models. This paper explores the one-pixel attacks in image classification, a black-box adversarial attack that introduces changes to the pixels of the input images to make the classifier predict erroneously. We use a pragmatic approach by employing different evolutionary algorithms - Differential Evolution, Genetic Algorithms, and Covariance Matrix Adaptation Evolution Strategy - to find and optimise these one-pixel attacks. We focus on understanding how these algorithms generate effective one-pixel attacks. The experimentation was carried out on the CIFAR-10 dataset, a widespread benchmark in image classification. The experimental results cover an analysis of the following aspects: fitness optimisation, number of evaluations to generate an adversarial attack, success rate, number of adversarial attacks found per image, solution space coverage and level of distortion done to the original image to generate the attack. Overall, the experimentation provided insights into the nuances of one-pixel attacks and compared three standard evolutionary algorithms, showcasing each algorithm's potential and evolutionary computation's ability to find solutions in this strict case of the adversarial attack.

12:35
Measuring Similarities in Model Structure of Metaheuristic Rule Set Learners

ABSTRACT. We present a way to measure similarity between sets of rules for regression tasks. This was identified to be an important but missing tool to investigate Metaheuristic Rule Set Learners (MRSLs), a class of algorithms that utilize metaheuristics such as Genetic Algorithms to solve learning tasks: The commonly-used predictive performance–based metrics such as mean absolute error do not capture most users' actual preferences when they choose these kinds of models since they typically aim for model interpretability (i.e. low number of rules, meaningful rule placement etc.) and not low error alone. The similarity measure is based on a form of metaheuristic-agnostic edit distance and is meant to be used in conjunction with a certain class of benchmark problems for analysing and improving an as-of-yet underresearched part of MRSL algorithms: The metaheuristic that optimizes the model's sturcture (i.e. the set of rule conditions). We discuss the measure's most important properties and demonstrate its applicability by performing experiments on the best-known MRSL, XCSF, comparing it with two non-metaheuristic Rule Set Learners, Decision Trees and Random Forests.

12:45
Evolutionary Multi-Objective Optimization of Large Language Model Prompts for Balancing Sentiments
PRESENTER: Jill Baumann

ABSTRACT. The advent of large language models (LLMs) such as Chat- GPT has garnered significant attention in various domains, owing to their remarkable performance and versatility. As the utilization of these models continues to rise, the importance of effective prompt engineering has come to the fore. Prompt optimization emerges as a crucial chal- lenge, given its direct impact on model performance and the extraction of relevant information. Recently, evolutionary algorithms have shown promise in addressing this issue, paving the way for novel optimization strategies. In this work, we propose EMO-Prompts, a multi-objective evolutionary approach specifically tailored for prompt optimization, us- ing sentiment analysis as a case study. We leverage sentiment analy- sis capabilities as our experimental objectives. Our results demonstrate that EMO-Prompts significantly enhances the quality and effectiveness of prompts, leading to improved performance and more reliable sentiment analysis outcomes.

12:55
Evolutionary Feature-Binning with Adaptive Burden Thresholding for Biomedical Risk Stratification

ABSTRACT. Multivariate associations including additivity, feature interactions, heterogeneous effects, and rare feature states can present significant obstacles in statistical and machine-learning analyses. These relationships can limit the detection capabilities of many analytical methodologies when predicting outcomes including risk stratification in biomedical survival analyses. Feature Inclusion Bin Evolver for Risk Stratification (FIBERS) was previously proposed using an evolutionary algorithm to discover groups (i.e. bins) of features wherein the burden of feature values automatically determined the risk strata of a given instance in right censored survival analysis. A key limitation of FIBERS is that it assumes a fixed threshold for feature burden in stratifying high vs. low risk, which restricts the flexibility of bin discovery. In the present work, we extend FIBERS to include different strategies for adaptive burden thresholding such that feature bins are discovered alongside the threshold that best separates risk strata. Preliminary comparative performance evaluation was conducted across simulated datasets with different underlying ideal burden thresholds yielding performance improvements over the original FIBERS algorithm. This algorithmic feasibility study lays the groundwork for ongoing application to the real-world problem of kidney graft failure risk stratification in dealing with the expected population heterogeneity including differences in race, ethnicity, and sex.

13:05
Robust Neural Architecture Search using Differential Evolution for Medical Images

ABSTRACT. Over the last decade, deep learning has become very popular thanks to its high performance and ability to learn diverse representations. It has been widely applied to multiple disciplines, including Computer Vision (CV), Natural Language Processing (NLP), medical image analysis, etc. Progression in these areas assists researchers building efficient deep learning algorithms for several challenging problems. However, these architectures still exhibit some limitations, such as their vulnerability against adversarial attacks. Recent studies have demonstrated that Convolutional Neural Network (CNN) architectures are sensitive to adversarial attacks with imperceptible permutations. Adversarial attacks on medical images may cause manipulated decisions and decrease the performance of the diagnosis system. The robustness of medical systems is crucial, as it assures an improved healthcare system and assists medical professionals to make significant decisions. Various studies have been performed to secure medical systems against adversarial attacks. However, they have used handcrafted architectures. In this study, we propose an evolutionary neural architecture search approach to produce robust architectures against adversarial attacks on medical images. The Differential Evolution (DE) algorithm is used as a search algorithm. Furthermore, an attention-based search space is used, which consists of five different attention layers and sixteen convolution and pooling operations. Experiments on multiple datasets from the MedMNIST benchmark demonstrated the effectiveness of the proposal in comparison with deep learning architectures and robust NAS approach.

11:00-13:35 Session 3C: EvoApplications 1 - Image Analysis, Signal Processing and Pattern Recognition
Location: Room C
11:00
Integrating Data Augmentation in Evolutionary Algorithms for Feature Selection: A Preliminary Study
PRESENTER: Emanuele Nardone

ABSTRACT. This study investigates the efficacy of data augmentation and feature selection in improving machine learning models across various tabular datasets in terms of sample and number of features. Our research unfolds in three phases, each adding complexity and depth to the analysis. Initially, we set a baseline through standard preprocessing and classification techniques, establishing a foundational understanding of each dataset's characteristics. The second phase introduces feature selection, where we employ Particle Swarm Optimization (PSO), Genetic Algorithm (GA), SelectKBest, and Recursive Feature Elimination with Cross-Validation (RFECV). A key aspect of our approach is the utilization of a specially designed fitness function in both PSO and GA, aimed at enhancing model accuracy while simultaneously reducing the number of features. The final segment of our study explores the integration of data augmentation into our models. By enriching the original datasets with synthetic data, we examine how different levels of augmentation affect model performance. The results from this phase are particularly insightful, indicating that strategic data augmentation, when combined with our feature selection approach, leads to notable improvements in accuracy. This research underscores the potential of combining data augmentation with intelligent feature selection techniques to achieve more efficient and accurate machine learning models.

11:25
Incremental growth on CPPN based optimization of biohybrid actuators.

ABSTRACT. One of the training methods of Artificial Neural Networks is Neuroevolution (NE) or the application of Evolutionary Optimization on the architecture and weights of networks to fit the target behaviour. In order to provide competitive results, three key concepts of the NE methods require more attention, i.e., the crossover operator, the niching capacity and the incremental growth of the solutions' complexity. Here we study an appropriate implementation of the incremental growth for an application of NE on Compositional Pattern Producing Networks (CPPNs) that encode the morphologies of biohybrid actuators. The target for these actuators is to enable the efficient angular movement of a drug-delivering catheter in order to reach difficult areas in the human body. As a result, the methods presented here can be a part of a modular software pipeline that will enable the automatic design of Biohybrid Machines (BHMs) for a variety of applications. The proposed initialization with minimal complexity of these networks resulted in faster computation for the predefined computational budget in terms of number of generations, notwithstanding that the emerged champions have achieved similar fitness values with the ones that emerged from the baseline method. Here, fitness was defined as the maximum deflection of the biohybrid actuator from its initial position after 10 seconds of simulated time on an open-source physics simulator. Since, the implementation of niching was already employed in the existing baseline version of the methodology, future work will focus on the application of crossover operators.

11:50
Cellular Genetic Algorithms for identifying variables in hybrid Gene Regulatory Networks

ABSTRACT. The hybrid modelling framework of gene regulatory networks (hGRNs) is a functional framework for studying biological systems, taking into account both the structural relationship between genes and the continuous time evolution of gene concentrations. The goal is to identify the variables of such a model, controlling the aggregated experimental observations. A recent study considered this task as a free optimisation problem and concluded that metaheuristics are well suited. The main drawback of this previous approach is that panmictic heuristics converge towards one basin of attraction in the search space, while biologists are interested in finding multiple satisfactory solutions. This paper investigates the problem of multimodality and assesses the effectiveness of cellular genetic algorithms (cGAs) in dealing with the increasing dimensionality and complexity of hGRN models. A comparison with the second variant of covariance matrix self-adaptation strategy with repelling subpopulations (RS-CMSA-ESII), the winner of the CEC'2020 competition for multimodal optimisation (MMO), is made. Results show evidence that cGAs better maintain a diverse set of solutions while giving better quality solutions, making them better suited for this MMO task.

12:15
Evolving Staff Training Schedules using an Extensible Fitness Function and a Domain Specific Language.

ABSTRACT. When using a meta-heuristic based optimiser in some industrial scenarios, there may be a need to amend the objective function as time progresses to encompass constraints that did not exist during the development phase of the software. We propose a means by which a Domain Specific Language (DSL) can be used to allow constraints to be expressed in language familiar to a domain expert, allowing additional constraints to be added to the objective function without the need to recompile the solver. To illustrate the approach, we consider the construction of staff training schedules within an organisation where staff are already managed within highly constrained schedules. A set of constraints are hard-coded into the objective function in a conventional manner as part of a Java application. A domain specific language (named Basil) is presented which is used to specify additional constraints affecting individual members of staff or groups. We demonstrate the use of Basil and show how it allows the specification of additional constraints that enable the software to meet the requirements of the user without any technical knowledge.

12:25
Integrating Bayesian and Evolutionary Multi-objective Optimisation

ABSTRACT. Both Multi-Objective Evolutionary Algorithms (MOEAs) and Multi-Objective Bayesian Optimisation (MOBO) are designed to address challenges posed by multi-objective optimisation problems. MOBO offers the distinct advantage of managing computationally or financially expensive evaluations by constructing Bayesian models based on the dataset. MOBO employs an acquisition function to strike a balance between convergence and diversity, facilitating the selection of an appropriate decision vector. MOEAs, similarly focused on achieving convergence and diversity, employ a selection criterion. This paper contributes to the field of multi-objective optimisation by constructing Bayesian models on the selection criterion of decomposition-based MOEAs within the framework of MOBO. The modelling process incorporates both mono and multi-surrogate approaches. The findings underscore the efficacy of MOEA selection criteria in the MOBO context, particularly when adopting the multi-surrogate approach. Evaluation results on both real-world and benchmark problems demonstrate the superiority of the multi-surrogate approach over its mono-surrogate counterpart for a given selection criterion. This study emphasises the significance of bridging the gap between these two optimisation fields and leveraging their respective strengths.

12:35
A Hierarchical Approach to Evolving Behaviour-Trees for Swarm Control

ABSTRACT. Behaviour trees (BTs) are commonly used as controllers in robotic swarms due their modular composition and to the fact that they can be easily interpreted by humans. From an algorithmic perspective, an additional advantage is that extra modules can easily be introduced and incorporated into new trees. Genetic Programming (GP) has already been shown to be capable of evolving BTs to achieve a variety of sub-tasks (primitives) of a higher-level goal. In this work we show that a hierarchical controller can be evolved that first uses GP to evolve a repertoire of primitives expressed as BTs, and then to evolve a high-level BT controller that leverages the evolved repertoire for a foraging task. We show that the hierarchical approach that uses BTs at two levels outperforms a baseline in which the BTs are evolved using only low-level nodes. In addition, we propose a method to improve the quality of the primitive repertoire, which in turn results in improved high-level BTs.

12:45
Finding sets of solutions for temporal uncertain problems

ABSTRACT. The multi-objective pathfinding problem is a complex and NP-hard problem with numerous industrial applications. However, the number of non-dominated solutions can often exceed human comprehension capacity. This paper introduces a novel methodology that leverages the concept of a Pareto graph to address this challenge. Unlike previous approaches, our method constructs a graph that relates paths where there is potential for change between them and applies a graph community algorithm to identify solution subsets based on specific aspects defined by a decision-maker. We describe the construction of a Route Change Graph (RCG) to represent possible route changes. A matrix is constructed to save the number of possible change opportunities between two routes, which is then used to construct the RCG. We propose using a threshold value for edge weights in the graph construction, balancing between minimising the number of edges and maintaining connectivity. Following the construction of the RCG, we apply a community detection algorithm to identify closely related solutions, using Leiden algorithm due to its efficiency and refinement phase. We propose calculating various metrics on these communities, including Density, Average Cluster Coefficient, Group Betweenness Centrality, and Graph Degree Centrality, to provide insights into the network structure and interconnectivity. This methodology offers a more manageable set of solutions for decision-makers, enhancing their ability to make informed decisions in complex multi-objective pathfinding problems.

12:55
Iterated Beam Search for Wildland Fire Suppression

ABSTRACT. Wildfires cause significant damage costs globally, and it is likely that they are becoming more damaging due to climate change. Here we study methods for fire suppression, after a breakout of fire. In our model, we have a grid graph $G=(V,A)$ that represents the discretization of a terrain into cells and an ignition node $s \in V$ from which the fire spreads to other nodes. The spread of the fire is defined by the arc weights, which can be used to model important factors such as wind direction and vegetation type. At various points in time, one or more fire suppression resources become available to be applied to nodes in the graph that are not yet burned. Applying a resource to a node $v \in V$ adds a delay to the outgoing edges of $v$, which causes a local slowdown in fire propagation. The goal is to find an allocation of resources to the nodes of the graph such that the total burned area at a target time is minimized. In this work, we propose a heuristic algorithm based on beam search to tackle this problem. Our computational experiments show that our approach is able to consistently find the optimal solution to almost all instances used in literature, but in considerably less time than previous approaches.

11:00-13:35 Session 3D: Late-Breaking Abstracts
Location: Room D
11:00
Two fast but unsuccessful algorithms \\for generating randomly folded proteins in HP

ABSTRACT. At the time of writing, there is no known deterministic time algorithm to uniformly randomly sample valid initial solutions for the HP protein folding model. In this short paper, we explore the difference between sampling prefabricated uniformly randomized conformations and runtime randomized sampling with collision avoidance. The last one performs better, at the cost of triple computation time, but is it truly uniform?

11:10
Frequency Fitness Assignment for Untangling Proteins in 2D

ABSTRACT. At the time of writing, there is no known deterministic-time algorithm to sample valid initial solutions with uniform random distribution for the HP protein folding model, because guaranteed uniform random sampling produces collisions (i.e. constraint violations). The expected number of collisions increases so fast with problem instance size that resampling is also infeasible. Maybe we can try uniform random sampling, and then hillclimb our way out of the violations? Here, we report results on a regular hillClimbers and an FFA-hillClimber, which traverses the search space following rare objective values instead of good objective values, applied to the task.

11:20
The Max-Cut’s Problem Solved with a Hybrid Evolutionary-QAOA Method

ABSTRACT. We present a method that solves the classical Max-Cut problem by combining an Evolutionary Algorithm (EA) with a Quantum Approximate Optimization Algorithm (QAOA). First, we compare the widely-used QAOA formulation using a Cobyla optimizer against our evolutionary-based approach to optimize the ansatz parameters. Then, we show that using an evolutionary strategy is competitive in achieving the same or better accuracy of solution with lower variance than Cobyla. In our experiments, the EA also requires fewer function evaluations to achieve the above results, increasing the problem size from 4 to 26 qubits. In the latter case, the EA reaches the optimal solution whilst Cobyla only reaches a solution that is on average roughly 80\% of the best Max-Cut.

11:30
Shape of the Waterfall: Solvability Transitions in the QAP

ABSTRACT. We consider a special formulation of the quadratic assignment problem (QAP): QAP-SAT, where the QAP is composed of smaller sub-problems or clauses which can be satisfied. A recent study showed a steep drop in solvability in relation to the number of clauses in QAP-SAT and robust taboo seach. In this work we characterise the nature of the solvability curves for this new class of QAP instances.

11:40
Migrolanding, the distribution of migrants as a solution to a reverse colonization problem

ABSTRACT. We define a new multi-objective optimization problem that consists in assigning migrants to different geographical subareas and, optionally, providing a job contract to them trying to maximize the satisfaction of the migrants and cover the labor needs of the different subareas. We provide an integer linear program for the problem and prove that the problem is NP-hard. We also point to potential sources of information to generate realistic instances for the problem.

11:50
Using large language models for iterative enhancement of metaheuristics: a viable path forward?

ABSTRACT. This research investigates the potential of leveraging advanced large language models (LLMs) for the iterative enhancement of metaheuristic algorithms. By repeatedly prompting an LLM, this study aims to achieve measurable improvements in algorithm performance. We argue that LLMs could, in the future, serve as a powerful tool for automated algorithm design.

12:00
Best practices for energy-thrifty evolutionary algorithms in the low-level language zig

ABSTRACT. The most fruitful way of making evolutionary algorithms spend the least amount of energy is to consider all possible programming techniques and platform choices that could, theoretically, affect performance, and carry out experiments using EA workloads in different platforms, eventually choosing those techniques that yield the minimum amount of energy expenses. These techniques include a choice of different data structures, as well as affecting compilation in such a way that energy footprint is reduced; they have to be replicated in different computing platforms because these expenditures may be affected by all the layers of the operating system and runtime framework used. In this paper we will experiment with different data structures and code refactoring techniques in the low-level language zig, trying to design rules of thumb that will help developers create green evolutionary algorithms. We will include two different hardware platforms, looking for the one that spends the least energy.

12:10
GP2SO: modeling and embedding 1-D signals using parametric symbolic regression

ABSTRACT. We briefly summarize the ideas underlying GP2SO, a hybrid GP/PSO-based method that produces a latent signal representation using parametric symbolic regression, discussing its possible applications and issues that might need to be tackled when applying the method to real-world data.

12:20
Symbiotic Connectivity: Optimizing Rural Digital Infrastructure with Solar-Powered Mesh Networks Using Multi-Objective Evolutionary Algorithms

ABSTRACT. I present an open-source, ecologically integrated model for rural connectivity, merging the location of nodes mesh networks with renewable energy systems. Employing evolutionary algorithms, this approach optimizes node placement for internet access and symbiotic energy distribution. This model, grounded in community collaboration, demonstrates a balance between technological advancement and environmental stewardship, offering a blueprint for sustainable infrastructure in similar rural settings.

12:30
The Evolution of Legibility: Exploring Type Design with an Evolutionary Generative Adversarial Model

ABSTRACT. Type design legibility is a critical aspect of design, influencing how easily viewers can comprehend written content. However, less legibility can lead to more expressive glyphs, and sometimes, that is exactly what we aim for. This research explores the evolutionary process of generating typefaces that optimize legibility using an Evolutionary Generative Adversarial Model. By using a generative adversarial model with an evolutionary generator, we are taking advantage of not only the final result but also the process of getting there, exploring the evolution of legibility. As an ongoing project, various possibilities are available, providing a new approach to generating type design and adjusting legibility levels according to specific design goals.

12:40
Combining Genetic Algorithms and Ant Colony Optimization for the Effective Allocation of Virtual Network Functions in a 5G Network

ABSTRACT. Devices with Internet access rise daily, which means also an increase of traffic data. In order to offer requested services, new and better techniques and tecnologies are required, such as NVF (Network Function Virtualization) in Software-Defined Networks (SDNs). In this paper we will present a method based on Genetic Algorithms to allocate virtualization of network functions (VNFs) along the different nodes of a network model, in order to maximize the performance of the service for users. To this end, an Ant Colony Optimization approach will be applied to evaluate the possible solutions of the GA, solving the so-called Service Function Chaining (SFC) problem. This will assign a fitness value to every possible individual of the algorithm.

14:15-16:05 Session 4A: EuroGP Best Paper Nominations
Location: Room A
14:15
Grammar-based Evolution of Polyominoes

ABSTRACT. Languages that describe two-dimensional (2-D) structures have emerged as powerful tools in various fields, encompassing pattern recognition and image processing, as well as modeling physical and chemical phenomena. One kind of two-dimensional structures is given by labeled polyominoes, i.e., geometric shapes composed of connected unit squares represented in a 2-D grid. In this paper, we present (a) a novel approach, based on grammars, for describing sets of labeled polyominoes that meet some predefined requirements and (b) an algorithm to develop labeled polyominoes using the grammar. We show that the two components can be used for solving optimization problems in the space of labeled polyominoes, similarly to what happens for strings in grammatical evolution (and its later variants). We characterize our algorithm for developing polyominoes in terms of representation-related metrics (namely, validity, redundancy, and locality), also by comparing different representations. We experimentally validate our proposal using a simple evolutionary algorithm on a few case studies where the goal is to obtain a target polyomino: we show that it is possible to enforce hard constraints in the search space of polyominoes, using a grammar, while performing the evolutionary search.

14:40
Naturally Interpretable Control Policies via Graph-based Genetic Programming

ABSTRACT. In most high-risk applications, interpretability is crucial for ensuring system safety and trust. However, existing research often relies on hard-to-understand, highly parameterized models, such as neural networks. In this paper, we focus on the problem of policy search in continuous observations and actions spaces. We leverage two graph-based Genetic Programming (GP) techniques—Cartesian Genetic Programming (CGP) and Linear Genetic Programming (LGP)—to develop effective yet interpretable control policies. Our experimental evaluation on eight continuous robotic control benchmarks shows competitive results compared to state-of-the-art Reinforcement Learning (RL) algorithms. Moreover, we find that graph-based GP tends towards small, interpretable graphs even when competitive with RL. By examining these graphs, we are able to explain the discovered policies, paving the way for trustworthy AI in the domain of continuous control.

15:05
Investigating Premature Convergence in Co-optimization of Morphology and Control in Evolved Virtual Soft Robots

ABSTRACT. Evolving virtual creatures is a field with a rich history and recently it has been getting more attention, especially in the soft robotics domain. The compliance of soft materials endows soft robots with complex behavior, but it also makes their design process unintuitive and in need of automated design. Despite the great interest, evolved virtual soft robots lack the complexity, and co-optimization of morphology and control remains a challenging problem. Prior work identifies and investigates a major issue with the co-optimization process -- fragile co-adaptation of brain and body resulting in premature convergence of morphology. In this work, we expand the investigation of this phenomenon by comparing learnable controllers with proprioceptive observations and fixed controllers without any observations, whereas in the latter case, we only have the optimization of the morphology. Our experiments in two morphology spaces and two environments that vary in complexity show, concrete examples of the existence of high-performing regions in the morphology space that are not able to be discovered during the co-optimization of the morphology and control, yet exist and are easily findable when optimizing morphologies alone. Thus this work clearly demonstrates and characterizes the challenges of optimizing morphology during co-optimization. Based on these results, we propose a new body-centric framework to think about the co-optimization problem which helps us understand the issue from a search perspective. We hope the insights we share with this work attract more attention to the problem and help us to enable efficient brain-body co-optimization.

15:30
Genetic Improvement of Last Level Cache

ABSTRACT. With increasing reliance on multi-core parallel computing performance is evermore dominated by interprocessor data communication typically provided by last level cache shared between CPUs. In an 8 core 3.6 GHz desktop, the Magpie parameter tuning genetic improvement (GI) system was able to reduce L3 cache access (load + stores) four fold on an existing open source 7000 line C PARSEC parallel computing VIPS image benchmark.

15:40
A Comprehensive Comparison of Lexicase-Based Selection Methods for Symbolic Regression Problems

ABSTRACT. Lexicase selection is a parent selection method that has been successfully used in many application domains. In recent years, several variants of lexicase selection have been proposed and analyzed. However, it is still unclear which lexicase variant performs best in the domain of symbolic regression. Therefore, we compare in this work relevant lexicase variants on a wide range of symbolic regression problems. We conduct experiments not only over a given evaluation budget but also over a given time as practitioners usually have limited time for solving their problems. Consequently, this work provides users a comprehensive guide for choosing the right selection method under different constraints in the domain of symbolic regression. Overall, we find that down-sampled epsilon-lexicase selection outperforms other selection methods on the studied benchmark problems for the given evaluation budget and for the given time. The improvements with respect to solution quality are up to 68% using down-sampled epsilon-lexicase selection given a time budget of 24h.

14:15-16:05 Session 4B: EvoMUSART 2 - Short Presentations
Location: Room B
14:15
Building an Embodied Musicking Dataset for co-creative music-making

ABSTRACT. In this paper, we present our findings of the design, development and deployment of a proof-of-concept dataset that captures some of the physiological, musicological, and psychological aspects of embodied musicking. After outlining the conceptual elements of this research, we explain the design of the dataset and the process of capturing the data. We then introduce two tests we used to evaluate the dataset: a) using data science techniques and b) a practice-based application in an AI-robot digital score. The results from these tests are conflicting: from a data science perspective the dataset could be considered garbage, but when applied to a real-world musicking situation performers reported it was transformative and felt to be ‘co-creative’. We discuss this duality and pose some important questions for future study. However, we feel that the datatset contains a set of relationships that are useful to explore in the creation of music.

14:25
Weighted Initialisation of Evolutionary Instrument and Pitch Detection in Polyphonic Music

ABSTRACT. Current state-of-the-art methods for instrument and pitch detection in polyphonic music often require large datasets and long training times; resources which are sparse in the field of music information retrieval, presenting a need for unsupervised alternative methods that do not require such prerequisites. We present a modification to an evolutionary algorithm for polyphonic music approximation through synthesis that uses spectral information to initialise populations with probable pitches. This algorithm can perform joint instrument and pitch detection on polyphonic music pieces without any of the aforementioned constraints. Sets of tuples of (instrument, style, pitch) are graded with a COSH distance fitness function and finally determine the algorithm's instrument and pitch labels for a given part of a music piece. Further investigation into this fitness function indicates that it tends to create false positives which may conceal the true potential of our modified approach. Regardless of that, our modification still shows significantly faster convergence speed and slightly improved pitch and instrument detection errors over the baseline algorithm on both single onset and full piece experiments.

14:35
Adaptation and Optimization of AugmentedNet for Roman Numeral Analysis Applied to Audio Signals

ABSTRACT. Automatic music harmony analysis has recently been significantly improved by AugmentedNet, a convolutional recurrent neural network for predicting Roman numeral labels. The original network receives perfect note annotations from the digital score as inputs and predicts various tonal descriptors: key, chord root, bass note, harmonic rhythm, etc. However, for many music tracks the score is not available at hand. For this study, we have first adjusted AugmentedNet for a direct application to audio signals represented either by chromagrams or semitone spectra. Second, we have implemented and compared further modifications to the network architecture: a preprocessing block designed to learn pitch spellings, increase of the network size, and addition of dropout layers. The statistical analysis helped to identify the best among all proposed configurations and has shown that some of the optimization steps significantly increased the classification performance. Besides, AugmentedNet can reach similar accuracies with audio features as inputs, compared to the perfect annotations that it was originally designed for.

14:45
Motifs, Phrases, and Beyond: The Modelling of Structure in Symbolic Music Generation

ABSTRACT. Modelling musical structure is vital yet challenging for artificial intelligence systems that generate symbolic music compositions. This literature review dissects the evolution of techniques for incorporating coherent structure, from early rule and template-based approaches to recent deep learning models. It provides foundational context by examining symbolic methods that imposed musical form using concepts from music theory. With the rise of deep learning, several RNN and Transformer based architectures have shown promise capturing repetitive patterns, hierarchical relationships and long-term dependencies. An emerging technique involves decomposing music generation into separate high-level structural planning and detailed content creation stages. Some systems incorporate musical knowledge by extracting melodic skeletons or templates to guide generation. Optimization, statistical, and pre-training techniques that learn general structural representations also hold potential. While progress is evident in capturing motifs and repetitions, modelling nuanced development of themes across extended compositions remains difficult. Key future directions include moving beyond rigid structural units, improving evaluation practices, and combining planning, pre-training, and hybrid neuro-symbolic approaches to fully realize their synergistic benefits. Mastering subtle long-term coherence akin to human composers remains elusive but may be attainable as complementary AI techniques continue to advance.

14:55
Generating emotional music based on improved C-RNN-GAN

ABSTRACT. This study introduces an emotion-based music generation model built upon the foundation of C-RNN-GAN, incorporating conditional GAN, and utilizing emotion labels to create diverse emotional music. Two evaluation methods were employed in this study to assess the quality and emotional expression of the generated music. Objective evaluation utilized metric calculations, comparing the generated music to the music in the EMOPIA database, including factors like note range, chord count, and chord consistency. Additionally, subjective assessment involved inviting 20 listeners to hear a set of both real and generated music. Listeners were asked to distinguish between real and generated music and evaluate emotional expression and harmony. The results indicate that the model-generated music successfully conveys a variety of emotions and approaches the quality of human-composed music

15:05
Modelling individual aesthetic preferences of 3D sculptures

ABSTRACT. Aesthetic preference is a complex puzzle with many subjective aspects. This subjectivity makes it incredibly difficult to computationally model aesthetic preference for an individual. Despite this complexity, individual aesthetic preference is an important part of life, impacting a multitude of aspects, including romantic and platonic relationships, decoration, product choices and artwork. Models of aesthetic preference form the basis of automated and semi-automated Evo-Art systems. These range from looking at individual aspects to more complex models considering multiple, different criteria. Effectively modelling aesthetic preference greatly increases the potential impact of these systems. This paper presents a flexible computational model of aesthetic preference, primarily focusing on generating 3D sculptures. Through demonstrating the model using several examples, it is shown that the model is flexible enough to identify and respond to individual aesthetic preferences, handling the subjectivity at the root of aesthetic preference and providing a good base for further extension to strengthen the ability of the system to model individual aesthetic preference.

15:15
The Forest: Towards Emergent Collaborative Art through Human Swarming

ABSTRACT. The Forest is a modular system of pillars inviting an audience to create an emergent audio-visual landscape through collective interactions over time. The pillars are designed to invoke reaction from participants through function and visual aesthetic, and to encourage collaboration – whether direct or indirect – towards an evolving piece of art to be experienced as a whole. In this paper, we show two variations of the Forest which we demonstrated at public outreach events. The first variation has no explicit interaction rules whereas in the second variation, we introduce “human swarm” rules towards emergent patterns in the installation. We show that the “human swarm” rules are able to produce diverse emergent outputs through simulation. Finally, we present anecdotal observations of the outcomes from outreach events, where interaction feedback empowers individuals in their creative exploration to produce consensus art.

15:25
Evolving User Interfaces: A Neuroevolution Approach for Natural Human-Machine Interaction

ABSTRACT. Intelligent user interfaces for human-machine interaction should be intuitive, invisible, and embodied in the user's natural physical environment. However, current computing systems still lack interfaces that fully capture complex visual and auditory perceptions.

In this study, we propose a neuroevolution approach, employing artificial neural networks optimized by evolutionary algorithms to evolve and accurately translate user inputs into system commands. This methodology adapts and refines system responses by evolving to accommodate diverse user inputs into precise system commands.

Our findings confirm the effectiveness of this approach, particularly in scenarios involving high-dimensional input spaces. The evolving interfaces developed herein show potential for improved user experiences and could pave the way for intelligent systems with natural user interfaces that understand and respond to user needs effectively.

The successful application of neuroevolution methods underscores their utility in creating natural, intuitive, and personalized interfaces. The study illustrates the potential of such approaches to revolutionize human-computer interaction by allowing organic and unobtrusive interfaces that adapt to individual user behaviors and preferences.

14:15-16:05 Session 4C: EvoApplications 2 - Problem Landscape Analysis & Other Applications
Location: Room C
14:15
Hilbert curves for efficient exploratory landscape analysis neighbourhood sampling

ABSTRACT. Landscape analysis aims to characterise optimisation problems based on their objective (or fitness) function landscape properties. The problem search space is typically sampled, and various landscape features are estimated based on the samples. One particularly salient set of features is information content, which requires the samples to be sequences of neighbouring solutions, such that the local relationships between consecutive sample points are preserved. Generating such spatially correlated samples that also provide good search space coverage is challenging. It is therefore common to first obtain an unordered sample with good search space coverage, and then apply an ordering algorithm such as the nearest neighbour to minimise the distance between consecutive points in the sample. However, the nearest neighbour algorithm becomes computationally prohibitive in higher dimensions, thus there is a need for more efficient alternatives. In this study, Hilbert space-filling curves are proposed as a method to efficiently obtain high-quality ordered samples. Hilbert curves are a special case of fractal curves, and guarantee uniform coverage of a bounded search space while providing a spatially correlated sample. We study the effectiveness of Hilbert curves as samplers, and discover that they are capable of extracting salient features at a fraction of the computational cost compared to Latin hypercube sampling with post-factum ordering. Further, we investigate the use of Hilbert curves as an ordering strategy, and find that they order the sample significantly faster than the nearest neighbour ordering, without sacrificing the saliency of the extracted features.

14:40
Predicting Algorithm Performance in Constrained Multiobjective Optimization: A Tough Nut to Crack

ABSTRACT. Predicting algorithm performance is crucial for selecting the best performing evolutionary algorithm for a given optimization problem. While some research on this topic has been done for single-objective optimization, algorithm performance prediction for constrained multiobjective optimization is still largely unexplored. In this work, we study two methodologies as candidates for predicting algorithm performance on 2D constrained multiobjective optimization problems. The first one consists of using state-of-the-art exploratory landscape analysis (ELA) features, designed specifically for constrained multiobjective optimization, as input to classical machine learning methods, and applying the resulting models to predict the performance classes. As an alternative methodology, we propose an end-to-end deep neural network trained to predict algorithm performance from a suitable problem representation, without relying on ELA features. The experimental results obtained on benchmark problems with three multiobjective optimizers show that neither of the two methodologies is capable of substantially outperforming a dummy classifier. This suggests that, with the current benchmark problems and ELA features, predicting algorithm performance in constrained multiobjective optimization remains a challenge.

15:05
On the Latent Structure of the bbob-biobj Test Suite

ABSTRACT. Landscape analysis is a popular method for the characterization of black-box optimization problems. It consists of a sequence of operations that, from a limited sample of solutions, approximate and describe the hypersurfaces formed by characteristic problem properties. The hypersurfaces, called problem landscapes, are described by sets of carefully crafted features that ought to capture their characteristic properties. In this way, arbitrary optimization problems with potentially very different technical parameters, such as search space dimensionality, are projected into specific feature spaces where they can be further studied. The representation of a problem in a feature space can be used, for example, to find similar problems and identify metaheuristic optimization algorithms that have the best track record on the same type of tasks. Because of that, the quality and properties of problem representation in the feature spaces gain importance. In this work, we study the representation properties of the popular bbob-biobj test suite in the space of bi-objective features, analyze the structure naturally emerging in the feature space, and analyze the high-level properties of the projection. The obtained results clearly demonstrate the discrepancies between the latent structure of the test suite and its expert perception.

15:30
A Novel Two-Level Clustering-Based Differential Evolution Algorithm for Training Neural Networks

ABSTRACT. Determining appropriate weights and biases for feed-forward neural networks is a critical task. Despite the prevalence of gradient-based methods for training, these approaches suffer from sensitivity to initial values and susceptibility to local optima. To address these challenges, we introduce a novel two-level clustering-based differential evolution approach, C2L-DE, to identify the initial seed for a gradient-based algorithm. In the initial phase, clustering is employed to detect various regions within the search space. Population updates are then executed based on the information available within each area. A new central point is proposed in the subsequent phase, leveraging cluster centres for incorporation into the population. Our C2L-DE algorithm is compared against several recent DE-based neural network training algorithms, and shown to yield favourable performance.

15:55
A New Angle: On Evolving Rotation Symmetric Boolean Functions

ABSTRACT. Rotation symmetric Boolean functions represent an interesting class of Boolean functions as they are relatively rare compared to general Boolean functions. At the same time, the functions in this class can have excellent properties, making them interesting for various practical applications. The usage of metaheuristics to construct rotation symmetric Boolean functions is a direction that has been explored for almost twenty years. Despite that, there are very few results considering evolutionary computations. This paper uses several evolutionary algorithms to evolve rotation symmetric Boolean functions with different properties. Despite using generic metaheuristics, we obtain results that are competitive with prior work relying on customized heuristics. Surprisingly, we find that bitstring and floating point encodings work better than the tree encoding. Moreover, evolving highly nonlinear general Boolean functions is easier than rotation symmetric ones.

16:25-17:45 Session 5A: EvoMUSART Best Paper Nominations
Location: Room A
16:25
Collaborative Interactive Evolution of Art in the Latent Space of Deep Generative Models

ABSTRACT. Generative Adversarial Networks (GANs) have shown great success in generating high quality images and are thus used as one of the main approaches to generate art images. However, usually the image generation process involves sampling from the latent space of the learned art representations, allowing little control over the output. In this work, we first employ GANs that are trained to produce creative images using an architecture known as Creative Adversarial Networks (CANs), then, we employ an evolutionary approach to navigate within the latent space of the models to discover images. We use automatic aesthetic and collaborative interactive human evaluation metrics to assess the generated images. In the human interactive evaluation case, we propose a collaborative evaluation based on the assessments of several participants. Furthermore, we also experiment with an intelligent mutation operator that aims to improve the quality of the images through local search based on an aesthetic measure. We evaluate the effectiveness of this approach by comparing the results produced by the automatic and collaborative interactive evolution. The results show that the proposed approach can generate highly attractive art images when the evolution is guided by collaborative human feedback.

16:50
MoodLoopGP: Generating Emotion-Conditioned Loop Tablature Music with Multi-Granular Features

ABSTRACT. Loopable music generation systems enable diverse applications, but they often lack controllability and customization capabilities. We argue that enhancing controllability can enrich these models, with emotion being a crucial aspect influencing listener experience. Hence, building upon LooperGP, a loopable tablature generation model, this paper explores endowing systems with control over conveyed emotions. To enable such conditional generation, we propose integrating musical knowledge by utilizing multi-granular musical features during model training and inference. Specifically, we incorporate song-level features (Emotion Labels, Tempo, and Mode) and bar-level features (Tonal Tension) together to guide emotional expression. Through algorithmic and human evaluations, we demonstrate the approach's effectiveness in producing music conveying target emotions of happiness and sadness. An ablation study is also conducted to clarify the contributing factors behind our approach's results.

17:15
Investigating the viability of Masked Language Modeling for symbolic music generation in abc-notation

ABSTRACT. The dominating paradigm for modeling sequences (e.g. text, music) with deep learning is the causal (left-to-right) approach, which consists in learning to predict tokens sequentially given those coming before it. Another paradigm is masked language modeling, which consists of learning to predict the masked tokens of a sequence in no specific order, given all non-masked tokens. The former is usually more appropriate for generation, but the latter can also be used and also allows editing. This paper investigates the viability of masked language modeling applied to Irish traditional music represented in abc notation -- a text-based format. We present our model, called {\em abcMLM}, which enables the user to arbitrarily edit tunes while retaining some generation capabilities. In fact, we found that generation using this paradigm is more challenging and required us to leverage additional information coming from the dataset, in the form of musical structure, to be on par with previous models.

16:25-17:45 Session 5C: EvoApplications 3 - Games
Location: Room C
16:25
Using Evolution and Deep Learning to Generate Diverse Intelligent Agents

ABSTRACT. Emergent behaviour arises from the interactions between individual components of a system, rather than being explicitly programmed or designed. The evolution of interesting emergent behaviour in intelligent agents is important in some application areas, for example, video game non-playable characters (NPCs). Being able to produce a diversity of high-quality opponents makes human players more engaged in games. In this research, we use genetic programming to evolve intelligent agents in a predator-prey simulation. The task is to evolve prey agents that capture the prey agents in the environment. A main goal, however, is to evolve agents that exhibit interesting and diverse behaviours, rather than the usual ones commonly evolved. First, we train a convolutional neural network to recognize main instances of ``generic'' prey behaviour, as recorded by an image trace of the prey's movement behaviour in the environment. A training set for 6 generic behaviours was used to train the CNN. A training accuracy of 98\% was obtained, as well as a validation performance of 90\%. Next, some experiments were performed that merge the CNN with fitness within the genetic programming system. In one experiment, the CNN's classification values are used as a ``diversity score'' which, when weighted with the fitness score, allow both agent quality and diversity to be considered. The result of this is that, with the appropriate weighting, high-quality, diverse solutions are obtained. In another experiment, we use the CNN classification scores to encourage the evolution of one of the known classes of trained behaviours. Results were that this trained behaviour was indeed more frequently evolved, compared to genetic programming runs using fitness alone. One conclusion is that machine learning techniques are a powerful tool to use for assisting the automated generation of diverse, high-quality intelligent agents.

16:50
Strategies for Evolving Diverse and Effective Behaviours in Pursuit Domains

ABSTRACT. Gamer engagement with computer opponents is an important aspect of computer games. Players will be bored if computer opponents are predictable, and the game will be monotonous. Computer opponents that are both challenging and exhibit interesting and novel behaviours are ideal. This research explores different strategies that encourage diverse emergent behaviours for evolved intelligent agents, while maintaining good performance with the task at hand. We consider the pursuit domain, which consists of a single predator agent and twenty prey agents. The predator’s controller is evolved through genetic programming, while the preys’ controllers are hand-crafted. The fitness of a solution is calculated as the number of prey captured. Inspired by Lehman and Stanley’s novelty search strategy, the fitness is combined with a diversity score, determined by combining four rudimentary behaviour measurements. We combine these basic scores using the many objective optimization strategy known as “sum of ranks”, which is proven to effectively balance a high number of conflicting objectives in optimization problems.We also examine different population diversity strategies, as well as different weighting schemes for combining fitness and diversity scores. After producing sets of solutions for the above experiments, we manually tabulate higher-level emergent behaviour observed in the evolved predators. The use of K-nearest neighbours (K=32) with population archive, combined with a fitness: diversity weighting of 50:50, gave the best results, as it effectively balanced good fitness performance and diverse emergent behaviour.

17:00
Vision Transformers for Computer Go

ABSTRACT. Motivated by transformers' success in diverse fields like language understanding and image analysis, our investigation explores their potential in the game of Go. Specifically, we focus on analyzing Transformers in Vision. Through a comprehensive examination of factors like prediction accuracy, win rates, memory, speed, size, and learning rate, we underscore the significant impact transformers can make in the game of Go. Notably, our findings reveal that transformers outperform the previous state-of-the-art models, demonstrating superior performance metrics. This comparative study was conducted against conventional Residual Networks.

17:10
Evolutionary Computation meets Stream Processing

ABSTRACT. Evolutionary computation has a great potential of exploiting parallelization, a feature often underemphasized when describing evolutionary algorithms (EAs). In this paper, we show that the paradigm of stream processing (SP) can be used to express EAs in a way that allows the immediate exploitation of parallel and distributed computing, not at the expense of the agnosticity of the EAs with respect to the application domain. We introduce the first formal framework for EC based on SP and describe several building blocks tailored to EC. Then, we experimentally validate our framework and show that (a) it can be used to express common EAs, (b) it scales when deployed on real-world stream processing engines (SPEs), and (c) it facilitates the design of EA modifications which would require a larger effort with traditional implementation.

17:35
Simple Efficient Evolutionary Ensemble Learning on Network Intrusion Detection Benchmarks

ABSTRACT. Training and deploying genetic programming (GP) models for intrusion detection tasks on the one hand remains a challenge (high cardinality and high class imbalance). On the other hand, GP solutions can also be particularly `lightweight' from a deployment perspective, enabling detectors to be deployed `at the edge' without specialized hardware support. We compare state-of-the-art ensemble learning solutions from GP and XGBoost on three examples of intrusion detection tasks with 250,000 to 700,000 training records, 8 to 115 features and 2 to 23 classes. XGBoost provides the most accurate solutions, but at two orders of magnitude higher complexity. Training time for the preferred GP ensemble is in the order of minutes, but the combination of simplicity and specificity is such that the resulting solutions are more informative and discriminatory. Thus, as the number of features increases and/or classes increase ensembles are composed from particularly simple trees that associate specific features with specific behaviours.

16:25-17:45 Session 5D: EuroGP 1
Chair:
16:25
SLIM_GSGP: The Non-Bloating Geometric Semantic Genetic Programming

ABSTRACT. Geometric semantic genetic programming (GSGP) is a successful variant of genetic programming (GP), able to induce a unimodal error surface for all supervised learning problems. However, a limitation of GSGP is its tendency to generate offspring significantly larger than their parents, resulting in continually growing program sizes. This leads to the creation of models that are often too complex for human comprehension. This paper presents a novel GSGP variant, the Semantic Learning algorithm with Inflate and deflate Mutations (SLIM). SLIM retains the essential theoretical characteristics of traditional GSGP, including the induction of a unimodal error surface. In addition to an inflate mutation, which produces larger offspring, SLIM introduces a novel geometric semantic mutation, the deflate mutation, which generates smaller offspring than its parents. The study introduces four SLIM variants and presents experimental results demonstrating that, across six symbolic regression test problems, SLIM consistently evolves models with equal or superior performance on unseen data compared to traditional GSGP and standard GP. These SLIM models are significantly smaller than those produced by traditional GSGP and are either smaller or of comparable size to standard GP models. Notably, the compactness of SLIM models allows for easier human interpretation.

16:50
Fuzzy Pattern Trees for Classification Problems Using Genetic Programming
PRESENTER: Allan de Lima

ABSTRACT. Interpretability has become an essential concern for Machine Learning (ML) models in critical areas such as healthcare, law and manufacturing industries. Fuzzy Pattern Trees (FPTs) are tree-based structures in which the internal nodes are fuzzy operators, and the leaves are fuzzy features. Representing input data using fuzzy logic usually brings more interpretability to the features since it separates the data into specific, often meaningful, parts of their domain, usually associated with a descriptive term. This work uses Genetic Programming (GP) to evolve FPTs and assess their performance on 20 benchmark classification problems. Furthermore, we experiment using Lexicase Selection with FPTs and demonstrate that selection methods based on aggregate fitness, such as Tournament Selection, produce more accurate models before analysing why this is the case. We also examine results from the perspective of the interpretability of the models, propose new parsimony pressure methods embedded in Lexicase Selection, and analyse their ability to reduce the size of the solutions. The results show that for most problems, at least one method could reduce the size significantly while keeping a similar accuracy.

17:15
Look into the Mirror: Evolving Self-Dual Bent Boolean Functions

ABSTRACT. Bent Boolean functions are important objects in cryptography and coding theory, and there are several general approaches for constructing such functions. Metaheuristics proved to be a strong choice as they can provide many bent functions, even when the size of the Boolean function is large (e.g., more than 20 inputs). While bent Boolean functions represent only a small part of all Boolean functions, there are several subclasses of bent functions providing specific properties and challenges. One of the most interesting subclasses comprises (anti-)self-dual bent Boolean functions.

This paper provides a detailed experimentation with evolutionary algorithms with the goal of evolving (anti-)self-dual bent Boolean functions. We experiment with two encodings and two fitness functions to directly evolve self-dual bent Boolean functions. Our experiments consider Boolean functions with sizes of up to 16 inputs, and we successfully construct self-dual bent functions for each dimension. Moreover, when comparing with the evolution of bent Boolean functions, we notice that the difficulty for evolutionary algorithms is rather similar. Finally, we also tried evolving secondary constructions for self-dual bent functions, but this direction provided no successful results.

17:25
An Algorithm Based on Grammatical Evolution for Discovering SHACL Constraints

ABSTRACT. The continuous evolution of heterogeneous RDF data has led to an increase of inconsistencies (missing data, errors, ...). We assume that these inconsistencies are inherent in RDF data graphs. SHACL is a W3C recommendation to express various constraints that the RDF data must conform: it allows the detection of violations from some nodes of the RDF graph against these constraints. However, acquiring representative and meaningful SHACL constraints from complex and very large RDF data graphs is very challenging and tedious. Consequently, several recent works focus on the automatic generation of these constraints. We propose an approach based on grammatical evolution (GE) for exploring representative SHACL constraints of an RDF data graph. The proposed algorithm exploits RDF data for the building and the assessment of candidate SHACL shapes. Finally, we select the best candidates through an evolutionary process. This approach uses a probabilistic SHACL validation to consider the inherent errors from RDF data. The results highlight the relevance of this approach in discovering SHACL shapes inspired by association rule patterns from a real-world RDF data graph.