Ensuring food security is one of today's most pressing societal challenges. Plant pathogens pose threats to this sustainability goal, causing substantial crop yield losses upon unrecognized infestation. Reliable early-stage pathogen detection thus is an integral subject in agricultural research. Spectral vegetation indices, such as the Normalized Difference Vegetation Index, have long been found to correlate with plant physiological processes and used for pathogen induced disease recognition. More recently, hyperspectral imaging combined with machine learning has attracted research attention. We propose an evolutionary learning method based on Genetic Programming variants as means to evolve tailored spectral vegetation indices, specific to pathogen-crop combinations and sensor modalities, from raw hyperspectral images, for which we provide a first proof-of-principle in the context of crop pathogen detection. To demonstrate our method's potential, we compare its capability to highlight symptomatic crop leaf areas at different infection stages against established indices, both quantitatively and qualitatively. Our empirical study results showcase the viability of the evolved tailored indices for recognizing Asian soybean rust, wheat yellow rust, and wheat powdery mildew on hyperspectral images of soybean and winter wheat leaves, with the evolved indices evaluating superior or at least comparable to established indices.

Deep learning–based weed control systems often struggle with limited training data diversity and constrained computational resources, restricting their effectiveness in real-world deployment. To address these limitations, we introduce a Stable Diffusion–based inpainting framework that progressively augments training datasets in 25% increments, up to 200%, enriching both data volume and variability. We systematically evaluate three state-of-the-art object detection architectures, such as large, small, and nano variants of YOLO11 and YOLOv12, along with large RT-DETR models, under three precision settings (FP32, FP16, INT8) using mAP50 and mAP50-95 evaluation metrics. Experiments on NVIDIA Jetson Orin Nano, NVIDIA Jetson AGX Orin, and spo-comm rugged computing unit reveal that quantization consistently reduces latency and memory footprint, with INT8 compression producing the most compact and fastest models. While INT8 often induces accuracy degradation, we show that this loss is significantly minimized by targeted synthetic augmentation. Notably, small YOLO variants trained with augmented data match, and in some cases surpass, the detection performance of their baseline large counterparts, without added model size or inference cost. Furthermore, utilizing the INT8-quantized Stable Diffusion for data generation preserves augmentation benefits on the downstream models while minimizing generation overhead. In combination, these contributions establish a novel training and deployment strategy for embedded AI in the context of weed detection, demonstrating that small YOLO models, INT8 quantization, and targeted synthetic augmentation can jointly deliver higher efficiency without sacrificing accuracy.

In automated crop protection tasks, deep learning has demonstrated significant potential. However, these advanced models rely heavily on high-quality, diverse datasets, which are often scarce and costly to obtain in agricultural settings. Traditional data augmentation techniques, while useful for increasing the volume of the dataset, often fail to capture the real-world variability needed for robust model training. In this paper, we present a novel method for generating synthetic images to enhance the training of deep learning-based object detection models for intelligent weed control, aiming to improve data efficiency. The architecture of our GenAI-based image generation pipeline integrates the Segment Anything Model (SAM) for zero-shot domain adaptation with a text-to-image Stable Diffusion Model, enabling the creation of synthetic images that can accurately reflect the idiosyncratic properties and appearances of a variety of real-world conditions. We further assess the application of these synthetic datasets on edge devices by evaluating state-of-the-art lightweight YOLO models, measuring data efficiency by comparing mAP50 and mAP50-95 scores among different proportions of real and synthetic training data. Incorporating these synthetic datasets into the training process has been found to result in notable improvements in terms of data efficiency. For instance, most YOLO models that are trained on a dataset consisting of 10% synthetic images and 90% real-world images typically demonstrate superior scores on mAP50 and mAP50-95 metrics compared to those trained solely on real-world images. The integration of this approach opens opportunities for achieving continual self-improvement of perception modules in intelligent technical systems.

Deep-Q-Networks are built in a way that, given a state, they predict the Q-values for the entire action-space. However, given an experience, the training update only incorporates the loss value for a single action-the one that has actually been executed. This is due to the rewards and follow-up states (required for computing the loss via the temporal-difference error) associated with the other actions being unknown. With these missing values at hand, or at least estimates of them, an update over the entire action-space would be possible. We present the Full-Update-DQN which is able to do just that. Sub-losses are weighted to compensate for uncertainty and noise and we are able to show in four different experiments in sparse reward settings, that our approach is able to solve these problems more consistently and even faster than the original approach.

Technological systems have been a part of human life since prehistory. Although they initially took the form of passive tools, such as axes and spoons, the Industrial Revolution saw the advent of powered, mechanized technology, operating “under it’s own steam,” without direct human control over every action. By integrating more complex information processing machinery, automation evolved into autonomy as decision-making and self-regulation became features of modern technology. Now, so-called intelligent systems, embodying techniques from the field of artificial intelligence (AI), are designed with the explicit intention of replicating rational behaviors and the sorts of things that minds do, inside technological systems.At the same time, the study of Artificial Life (ALife) (Langton, 1987) has explored the properties of living systems, both as they are found in nature, as they might be, and as humans can build them. This has exposed a large variety of mechanisms that produce qualities typically associated with life. Examples include self-organization, homeostasis, self-replication, evolution, learning, self-awareness, and many others besides.The Lifelike Computing Systems initiative (Stein et al., 2021b) aims to learn from the study of life and living systems to develop new, useful, “lifelike” systems; a further aim is to identify when such features are of value. The focus of this research direction is primarily on engineered technological systems broadly within the domain of computing.The notion of “lifelike computing” is not intended to separate itself from or replace previous initiatives; in a large number of cases, there are already technologies and research efforts that strongly lean toward lifelike computing systems in specific aspects. Building on a long and highly successful tradition in biologically inspired computing, the “lifelike” vision not only seeks inspiration in the living world but also seeks to replicate its qualities explicitly in technological systems. Indeed, we cannot claim that all bio-inspired systems remain lifelike, nor is this in general even always a desirable outcome for those designing bio-inspired systems. The agenda also goes beyond fundamental ALife research, often rightly exploratory in nature, because it focuses explicitly on building purposeful and reliable technological systems for people, based on ALife principles. Therefore the vision of explicit replication of lifelike qualities in technological systems of value to humanity marks a sharpening of focus.This special issue is a follow-up to the workshop series “Lifelike Computing Systems,” held at the International Conference on Artificial Life in 2020 and 2021 (Stein et al., 2021a), and again in 2022 (Stein et al., 2023). The workshop series hosted diverse talks showcasing early-stage research and work in progress, with topics ranging from plasticity in technical systems to artificial DNA, from self-explaining systems to realistic humanoid and animal robots.We have therefore solicited papers that explore and contribute to the discussion on research questions we deem key to be further explored: Which qualities of life are of high relevance and benefit for the engineering of lifelike computing systems useful to people? Why? How?How can we integrate and combine insights and methodological approaches from existing, related research initiatives, such as cybernetics, self-aware computing, organic computing, and autonomic computing?Which methods from domains like artificial life, bio-inspired computing, artificial intelligence, and self-adaptive and self-organizing systems contribute to achieving lifelike features of computing systems?When is more “lifelike” technology appropriate? What are the challenges associated with embedding technology that is more “lifelike” in society? How can these be tackled?This special issue represents an opportunity for more mature work emerging from this line of research to be presented. It contains four papers that together provide a review, analysis, and critique of the integration of lifelike properties into engineered systems, in many cases proposing concrete recommendations for future research directions and methods.In “Lessons from the Evolutionary Computation Bestiary,” Campelo and Aranha explore and critique the explosion of metaphor-centered metaheuristic methods that have been published in recent years and that claim to be inspired by—in their view—increasingly absurd natural phenomena. Examples surveyed include several different types of birds, mammals, fish, and invertebrates; soccer and volleyball; and even reincarnation, zombies, and gods. The authors acknowledge that metaphors can be powerful inspiration and explanatory tools and that, indeed, the field of metaheuristics has a long history of finding inspiration in natural systems, starting from evolution strategies, genetic algorithms, and ant colony optimization. However, they question the value of the emergence of hundreds of highly similar variants of essentially the same algorithm under different labels. The authors have curated a “bestiary” of such variants over the years, and their article in this issue reviews this, arguing that this proliferation has been counterproductive to scientific progress in the field. They argue that it does little to improve our ability to understand and simulate biological systems and that it can actively impede an improved understanding of how to design and analyze global optimization techniques. The article discusses why this social phenomenon in research may have occurred in recent years and its negative consequences, ending with a call to improve the scientific soundness of metaheuristic research.In “Does the Field of Nature-Inspired Computing Contribute to Achieving Lifelike Features?,” Tzanetos asks whether all nature-inspired algorithms remain lifelike. The article considers the history of evolutionary computation and, as in the first article, the proliferation of many so-called nature-inspired techniques in recent years. The author juxtaposes the value of such techniques in solving hard problems with an analysis to support an argument that the mathematics of these techniques often does not match the source behavior faithfully. In these cases, can it be said that the algorithms are indeed “lifelike,” and if not, does that matter, so long as they provide value in terms of their ability to solve problems intelligently? The article argues that historically, there was greater alignment between the algorithmic models and source behaviors, but this is often not seen in more recent attempts. The article ends by discussing if there is a need for new lifelike features of algorithms, concluding that this is not helpful—instead presenting recommendations for future research in nature-inspired computing, which, the authors argue, would move the field in “the right direction.”In “Assessing Model Requirements for Explainable AI: A Template and Exemplary Case Study,” Heider et al. explore the explainability of decision support systems that use evolutionary rule-based machine learning techniques, more precisely, learning classifier systems (LCSs). Self-adaptive and self-optimizing systems are necessarily dynamic, yet for them to be accepted by people in sociotechnical settings, explanations for machine-made decisions are often essential. The authors argue that rule-based machine learning models, such as LCSs, present an opportunity for transparent machine learning models that naturally support access to explanations. To assist with designing and evaluating such models, they also propose a generic and thus broadly applicable questionnaire template. The template is demonstrated to provide valuable insights for the design of such LCS models in specific scenarios. The approach is illustrated in a manufacturing case study.Finally, in “Artificial Collective Intelligence Engineering: A Survey of Concepts and Perspectives,” Casadei surveys computational techniques based on or harnessing “collectiveness,” often seen in many living systems, to produce capabilities beyond what can be achieved with individual or monolithic systems. A key concept common to these techniques is that such systems can exploit a large number of individuals to produce intelligent collective behavior out of not-so-intelligent components. The article argues that there is a trend in some areas of engineering toward this way of designing technological systems, citing examples such as the Internet of Things, swarm robotics, and crowd computing and emphasizing that these technologies span many techniques, systems, and application areas. An essential finding of the review is that there is substantial fragmentation of this research, however, and that the so-called “verticality” of research communities makes a common fundamental understanding of such systems challenging to achieve. The author argues that an important challenge is identifying, placing in a common structure, and ultimately connecting the different areas and methods addressing intelligent collectives. As such, the article presents a set of questions aimed at mapping out collective intelligence research. It uses this to develop a set of preliminary notions, concepts, and perspectives, as well as associated research opportunities, to develop a more fundamental understanding of computational collective intelligence engineering.The guest editors thank the authors of papers submitted to the “Lifelike Computing Systems” special issue as well as the reviewers, who gave valuable feedback to all the authors. We would also like to thank the organizers of the ALife conferences that hosted the Lifelike Computing Systems workshops as well as all the speakers and participants who contributed to many vibrant debates that informed the direction of the final set of articles in this issue. Last, we thank the Board of Editors of Artificial Life for supporting this special issue.

The concept of Experience Replay is a crucial element in Deep Reinforcement Learning algorithms of the DQN family. The basic approach reuses stored experiences to, amongst other reasons, overcome the problem of catastrophic forgetting and as a result stabilize learning. However, only experiences that the learner observed in the past are used for updates. We anticipate that these experiences posses additional valuable information about the underlying problem that just needs to be extracted in the right way. To achieve this, we present the Interpolated Experience Replay technique that leverages stored experiences to create new, synthetic ones by means of interpolation. A previous proposed concept for discrete-state environments is extended to work in continuous problem spaces. We evaluate our approach on the MountainCar benchmark environment and demonstrate its promising potential.

Abstract In the last decades, the classical Vehicle Routing Problem (VRP), i.e., assigning a set of orders to vehicles and planning their routes has been intensively researched. As only the assignment of order to vehicles and their routes is already an NP-complete problem, the application of these algorithms in practice often fails to take into account the constraints and restrictions that apply in real-world applications, the so called rich VRP (rVRP) and are limited to single aspects. In this work, we incorporate the main relevant real-world constraints and requirements. We propose a two-stage strategy and a Timeline algorithm for time windows and pause times, and apply a Genetic Algorithm (GA) and Ant Colony Optimization (ACO) individually to the problem to find optimal solutions. Our evaluation of eight different problem instances against four state-of-the-art algorithms shows that our approach handles all given constraints in a reasonable time.

Touch interfaces are human machine interface (HMI) that can be found in a wide range of products ranging from mobile phones over cars to home appliances.Many of these HMIs measure digital signals which are used to detect touch events.These signals are processed using filters in order to decide whether there is a touch event or not.The filterchain must be functional even if the signal contains heavy noise.Thus a precise calibration of the individual filters is necessary.We employ a genetic algorithm (GA) to choose the filter parameters automatically.We evaluate our approach in a series of experiments which includes simulated as well as real data.We additionally compare our GA with manually calibrated parameters and thereby show the superiority of our method in terms of the accuracy of the calibration provided.A cost-intensive manual calibration can thus be avoided.

Human machine interfaces (HMI) have become a part of our daily lives.They are an essential part of a variety of products ranging from computers over smart phones to home appliances.Customer's requirements for HMIs are rising and so does the complexity of the devices.Several years ago, many products had a rather simple HMI such as mere buttons.Nowadays lots of devices have screens that display complex text messages and a variety of objects such as icons.This leads to new challenges in testing, the goal of which it is to ensure quality and to find errors.We combine a genetic algorithm with computer vision techniques in order to solve two testing use cases located in the automated verification of displays.Our method has a low runtime and can be used on low budget equipment such as Raspberry Pi which reduces the operational cost in practice.

The research initiative “self-improving system integration” (SISSY) was established with the goal to master the ever-changing demands of system organisation in the presence of autonomous subsystems, evolving architectures, and highly-dynamic open environments. It aims to move integration-related decisions from design-time to run-time, implying a further shift of expertise and responsibility from human engineers to autonomous systems . This introduces a qualitative shift from existing self-adaptive and self-organising systems, moving from self-adaptation based on predefined variation types, towards more open contexts involving novel autonomous subsystems, collaborative behaviours, and emerging goals. In this article, we revisit existing SISSY research efforts and establish a corresponding terminology focusing on how SISSY relates to the broad field of integration sciences. We then investigate SISSY-related research efforts and derive a taxonomy of SISSY technology. This is concluded by establishing a research road-map for developing operational self-improving self-integrating systems.

XCS constitutes the most deeply investigated classifier system today. It offers strong potentials and comes with inherent capabilities for mastering a variety of different learning tasks. Besides outstanding successes in various classification and regression tasks, XCS also proved very effective in certain multi-step environments from the domain of reinforcement learning. Especially in the latter domain, recent advances have been mainly driven by algorithms which model their policies based on deep neural networks, among which the Deep-Q-Network (DQN) being a prominent representative. Experience Replay (ER) constitutes one of the crucial factors for the DQN's successes, since it facilitates stabilized training of the neural network-based Q-function approximators. Surprisingly, XCS barely takes advantage of similar mechanisms that leverage remembered raw experiences. To bridge this gap, this paper investigates the benefits of extending XCS with ER. We demonstrate that for single-step tasks ER yields strong improvements in terms of sample efficiency. On the downside, however, we reveal that ER might further aggravate well-studied issues not yet solved for XCS when applied to sequential decision problems demanding for long-action-chains.

An important component of many Deep Reinforcement Learning algorithms is the Experience Replay which serves as a storage mechanism or memory of made experiences. These experiences are used for training and help the agent to stably find the perfect trajectory through the problem space. The classic Experience Replay however makes only use of the experiences it actually made, but the stored samples bear great potential in form of knowledge about the problem that can be extracted. We present an algorithm that creates synthetic experiences in a nondeterministic discrete environment to assist the learner. The Interpolated Experience Replay is evaluated on the FrozenLake environment and we show that it can support the agent to learn faster and even better than the classic version.

Testing is a crucial part in the development of a new product.Due to the change from manual testing to automated testing, companies can rely on a higher number of tests.There are certain cases such as smoke tests where the execution of all tests is not feasible and a smaller test suite of critical test cases is necessary.This prioritization problem has just gotten into the focus of reinforcement learning.A neural network and an XCS classifier system have been applied to this task.Another evolutionary machine learning approach is the XCSF which produces, unlike XCS, continuous outputs.In this work we show that XCSF is superior to both the neural network and XCS for this problem.

The minimum set cover problem (MSCP) is one of the first NP-hard optimization problems discovered.Theoretically it has a bad worst case approximation ratio.As the MSCP turns out to appear in several real world problems, various approaches exist where evolutionary algorithms and metaheuristics are utilized in order to achieve good average case results.This work is intended to revisit and compare current results regarding the application of metaheuristics for the MSCP.Therefore, a recapitulation of the MSCP and its classification into the class of NP-hard optimization problems are provided first.After an overview of notable approximation methods, the focus is shifted towards a brief review of existing metaheuristics which were adapted for the MSCP.In order to allow for a targeted comparison of the existing algorithms, the theoretical worst case complexities in terms of the big O-notation are derived first.This is followed by an empirical study where the identified metaheuristics are examined.Here we use Steiner triple systems, Beasley's OR library, and introduce a new class of instances.Several of the considered approaches achieve close to optimal results.However, our analysis reveals significant differences in terms of runtime and shows that some approaches may even have exponential runtime.

Learning Classifier Systems (LCS) have been strongly investigated in the context of regression tasks and great successes have been achieved by applying the function approximating Extended Classifier System (XCSF) endowed with sophisticated prediction models. In this paper, a novel approach to model a classifier's payoff prediction is proposed. Radial Basis Function (RBF) interpolation is utilized as a new means to capture the underlying function surface complexity. We pose the hypothesis that by the use of a more flexible RBF-based classifier prediction, that alleviates the a priori bias injected via choosing the degree of a polynomial approximation, the classifiers can evolve toward a higher generality by maintaining at least a competitive level of performance compared to the current and probably mostly used state of the art approach - polynomial approximation in combination with the Recursive Least Squares (RLS) technique for incremental coefficient optimization. The presented experimental results underpin our assumptions by revealing that the RBF-based classifier prediction outperforms the n-th order polynomial approximation on several test functions of varying complexity. Additionally, results of experiments with various degrees of noise will be reported to touch upon the proposed approach's applicability in real world situations.

Carbon fiber reinforced plastics (CFRP) play a key role for the production of leightweight structures. Simultaneously, online quality inspection of CFRP becomes more important, especially for environments with high safety standards. In this context, vision systems aim to find defects of different shape, size, contour and orientation. Little effort, however, has been made in detecting defect areas in images taken from the surface of carbon fibers. A common approach for segmenting filament defects are edge detection and thresholding. With every change of material and process adjustments, the filter parameters have to be adapted. In this paper, we propose a cartesian genetic programming (CGP) approach to semi-automatically select the best parameters. This strategy saves time for parameter identification while at the same time increases precision. A test run on randomly selected samples shows how the approach can substantially improve detection reliability.

In this paper, we show how an evolutionary rule-based machine learning technique can be applied to tackle the task of self-configuration of smart camera networks.More precisely, the Extended Classifier System (XCS) is utilized to learn a configuration strategy for the pan, tilt, and zoom of smart cameras.Thereby, we extend our previous approach, which is based on Q-Learning, by harnessing the generalization capability of Learning Classifier Systems (LCS), i.e. avoiding to separately approximate the quality of each possible (re-)configuration (action) in reaction to a certain situation (state).Instead, situations in which the same reconfiguration is adequate are grouped to one single rule.We demonstrate that our XCS-based approach outperforms the Q-learning method on the basis of empirical evaluations on scenarios of different severity.

XCSF is a rule-based on-line learning system that makes use of local learning concepts in conjunction with gradient-based approximation techniques. It is mainly used to learn functions, or rather regression problems, by means of dividing the problem space into smaller subspaces and approximate the function values linearly therein. In this paper, we show how local interpolation can be incorporated to improve the approximation speed and thus to decrease the system error. We describe how a novel interpolation component integrates into the algorithmic structure of XCSF and thereby augments the well-established separation into the performance, discovery and reinforcement component. To underpin the validity of our approach, we present and discuss results from experiments on three test functions of different complexity, i.e. we show that by means of the proposed strategies for integrating the locally interpolated values, the overall performance of XCSF can be improved.

Distributed self-organising systems often face conflicts if more than one entity tries to access a limited resource. In order to solve this conflict, research focuses on techniques for resource allocation considering different priorities. In this paper, we propose to tackle the decision problem of whom to assign the resource by means of a co-evolutionary approach. We investigate appropriate fitness estimations, representation schemes, and configuration of the underlying genetic operators. We demonstrate the convergence and efficiency of our approach using an exemplary system model.