Kevin Mallinger

Software has supported forest management decisions since the 1970s, from growth simulators to reinforcement learning. Since 2023, large language models (LLMs) have been combined with tools, retrieval, memory and planning to form artificial intelligence (AI) agents that analyse forest data, decide on a sequence of steps, and either advise a person or act through software and machines. This survey reviews agentic AI in forestry. It traces the development of agent concepts in the field, describes the components that current agentic systems share, and reviews applications in seven areas, from forest inventory to biodiversity monitoring. We found only 13 LLM-based agentic systems in forestry. Most are single LLM agents or small groups of role-based agents that coordinate perception models, geospatial databases and retrieval over documents. Their evaluation relies on case studies, small expert panels and public datasets, and none of the reviewed LLM-based agents controls forest machines in the field. Recurring problems are factual errors, weak spatial reasoning, missing connectivity in forests, the cybersecurity of connected machines and trust, and current evaluations are short compared with the decades over which forest decisions take effect. We derive research directions on evaluation with forest simulators, edge deployment, interoperability, graded human oversight and safety constraints for forest machines. A comparison with agriculture identifies designs that can be transferred, and we expect agents to become routine first in analysis and advice, then in planning with forest models, and only later in the supervision of forest machines.

Random numbers are used to generate cryptographic keys, to run Monte Carlo simulations, and to train and sample machine learning models. Most software obtains them from deterministic pseudo-random number generators, which have good statistical properties but are predictable once their internal state is known. Quantum random number generators take their output from measurements on quantum systems, whose outcomes are not fixed by any prior state of the device. This article introduces the types of random number generators, reviews studies that compare quantum and pseudo-random numbers in numerical experiments, and lists the sources of randomness available in current computing systems. It then describes applications of quantum randomness in security, hashing, software, artificial intelligence, and public randomness. The quality of encryption is treated as depending on three factors: the algorithm, its implementation, and the randomness of keys and other security-relevant values. We give arguments for which applications benefit from quantum randomness, name further areas in which it could be tested, and outline next steps.

Fractional optimization methods and fractal activation functions are two independent directions for improving neural network training. Fractional optimizers extend first-order optimization through fractional derivatives and memory effects, whereas fractal activations introduce multi-scale nonlinear representations based on self-similar Weierstrass- and Blancmange-type functions. Here, we investigate their interaction within a unified experimental framework. We evaluate fractional optimizer families on Ackley and Himmelblau benchmark surfaces, in standard form and with additive Weierstrass-type perturbations, and then in feed-forward neural networks with conventional and fractal activations on ten classification datasets. The comparison includes standard methods, Herrera-type local scaling optimizers, explicit and adaptive memory-based fractional optimizers, and other representative literature methods. Fractional optimization and fractal activations show useful but selective pairings. Herrera-type local fractional scaling performs well with selected fractal activations in network training, while explicit Grünwald–Letnikov memory is competitive on the controlled surfaces without being preferentially favoured by the perturbation: the unmodified baselines rank first on both additive variants. Adaptive memory improves plain memory substitution in several cases, supporting controlled fractional memory as a promising direction rather than a universal replacement.

Fractional optimization methods and fractal activation functions have recently emerged as two independent directions for improving neural network training. Fractional optimizers extend first-order optimization through fractional derivatives and memory effects, whereas fractal activations introduce multi-scale nonlinear representations based on self-similar Weierstrass- and Blancmange-type functions. Despite their common mathematical motivation, both approaches have largely been studied independently. Here, we investigate their interaction within a unified experimental framework. We evaluate fractional optimizer families on Ackley and Himmelblau benchmark surfaces, in standard form and with additive Weierstrass-type perturbations, and then in feed-forward neural networks with conventional and fractal activations on ten OpenML classification datasets. The comparison includes standard methods, Herrera-style optimizers, explicit and adaptive memory-based fractional optimizers, and representative literature methods. Results are analysed by accuracy, convergence, cost, robustness, and optimizer–activation interaction. Overall, fractional optimization and fractal activations show useful but selective pairings. Herrera-style fractional scaling performs well with selected fractal activations in network training, while Grünwald–Letnikov memory is most relevant on perturbed surfaces. Adaptive memory improves plain memory substitution in several cases, supporting controlled fractional memory as a promising direction rather than a universal replacement.

This paper studies neural network layers that use learnable Lie-group actions as structured feature-space transformations. Instead of treating Lie groups only as input-domain symmetry constraints, the proposed approach embeds real-valued features into local vector banks, learns coordinates in a Lie algebra, maps these coordinates to group elements through the matrix exponential, and applies the resulting matrices to intermediate feature vectors. The framework supports groups such as SO(3), SU(2), and SU(3), and can be used either as a standalone structured backbone or as a component inside conventional neural architectures. The experimental evaluation covers several settings: tabular classification, tabular regression, synthetic signal denoising, generative adversarial learning, and recursive time-series forecasting. The classification and regression studies compare dense neural baselines, MLP–Lie hybrids, deeper Lie-group architectures, CatBoost, and ExtraTrees across repeated train-validation-test splits. The denoising experiment compares a classical autoencoder with an SU(3)-based autoencoder on synthetic oscillatory signals. The GAN experiment inserts an SU(3) layer into the discriminator and compares it with a standard convolutional GAN on MNIST digit generation. The time-series experiments compare a regular Transformer, a hybrid Transformer with one Lie-group layer, a Lie-group Transformer, and CatBoost under recursive holdout forecasting. The results show that Lie-group feature transformations are useful in selected settings, but they are not uniformly superior across all tasks. In classification, the structured models improve over the dense baseline on several datasets, while tree-based methods remain strongest on others. In regression, MLP–Lie models are competitive on some tasks, but CatBoost and ExtraTrees are often stronger. The clearest improvement is observed in signal denoising, where the Lie-group autoencoder reduces reconstruction error and improves signal-to-noise ratio. In the GAN experiment, the Lie-group discriminator gives moderate improvements in stability and discriminator metrics. In time-series forecasting, Lie-group Transformer variants improve over the regular Transformer on some series, while CatBoost remains a strong rolling-window baseline. Overall, the results support a dataset-dependent interpretation. Lie-group layers can act as useful structured feature mixers, especially when local vector structure or oscillatory behavior is relevant. At the same time, their benefit depends on the task, architecture, and computational cost. The framework therefore provides a practical basis for studying when algebraic feature-space transformations improve learning and when simpler baselines are sufficient.

Background: Forest fires are becoming more frequent and destructive under climate change, threatening ecosystems, and society. Existing machine learning approaches for modeling forest fire ignition vulnerability often lack the spatial resolution needed for local-scale estimation, rely on limited proxies for anthropogenic ignition pressure, and provide only limited insight into how ignition drivers and their importance vary across local conditions, which limits their usefulness for developing locally meaningful fire danger models. Results: We developed a machine learning framework for anthropogenic forest fire ignition vulnerability in Austria using 1569 documented fire events from 2012 to 2021 and an equal number of generated non-fire samples. The model integrates high-resolution environmental, socioeconomic, infrastructure-related, and human movement variables to estimate probabilistic vulnerability scores at spatially resolved point locations across Austria. A web based tool visualizes these predictions on a 50 by 50 m grid with adaptive aggregation at lower zoom levels. CatBoost and Random Forest achieved a precision of 80.7% and an accuracy of 77.9%, exceeding earlier Austrian studies with accuracies of around 60%. Both models produced highly consistent spatial patterns and identified fire prone areas especially in eastern and southern Austria. Explainable AI analyses showed positive effects of Running Frequency, Building Density, and Fine Fuel Moisture Code, while slope showed a negative effect, thereby making local ignition drivers and their spatial relevance transparent. Conclusions: The proposed framework enables high-resolution mapping of anthropogenic forest fire ignition vulnerability and Explainable Artificial Intelligence (XAI)-driven multi-criteria evaluation, while revealing the relative influence of environmental and anthropogenic drivers. The introduced models identify historically fire-prone zones in Austria at both local and regional scales and indicate that anthropogenic factors, together with established environmental conditions, are key drivers of ignition vulnerability. More broadly, the methodological approach and findings illustrate how interpretable and explainable machine learning can help assess which variables are relevant under different local conditions, thereby supporting targeted mitigation and adaptation planning as well as more context-aware model development. Further work is needed to evaluate ignition danger under changing anthropogenic and environmental pressures.

Artificial intelligence (AI) systems are increasingly proposed for agricultural decision support, precision farming and precision livestock farming (PLF), yet large-scale and durable adoption by practitioners remains the exception rather than the rule. Established maturity scales such as the Technology Readiness Level (TRL) capture the engineering progression of a technology from basic principles to operational deployment, but they are agnostic to whether end users actually rely on the resulting system in their daily work. This paper introduces the Technology Acceptance Level (TAL) as a complementary, human-centered metric for evaluating the usability, trustworthiness and operational uptake of AI in agriculture. TAL is defined as a nine-level ordinal scale, anchored in the Technology Acceptance Model and the European Union ethics guidelines for Trustworthy AI, that captures the trajectory from initial exposure of a system to a practitioner up to its stable, autonomous integration into routine decision-making. Each level is operationalised by a small set of measurable criteria spanning perceived usefulness, perceived ease of use, transparency, reliability under changing conditions, and alignment with agricultural workflows. We provide a mathematically grounded aggregation rule based on a non-compensatory, minimum-of-dimensions principle that prevents a single weak dimension from being masked by strong scores elsewhere, and we derive a confidence interval for TAL estimates from practitioner survey data. The framework is illustrated through three representative use cases—variable-rate nitrogen management in arable cropping, vision-based fruit detection for selective harvesting, and computer-vision-based lameness detection in dairy cattle—demonstrating how the same AI artefact can occupy very different acceptance levels across deployment contexts. By explicitly addressing the human-in-the-loop, TAL provides a practical instrument to diagnose adoption barriers, prioritise iterative system improvement, and steer the transition from technically functional prototypes to practically accepted, trustworthy AI in the agri-food sector.

Noise contamination is a common challenge in the analysis of time series data, where stochastic perturbations can obscure deterministic dynamics and complicate the interpretation of signals from chaotic and physiological systems. Reliable identification of noise regimes and their intensity is therefore essential for robust analysis of dynamical and biomedical signals, where incorrect attribution of stochastic perturbations can lead to misleading interpretations of system behavior. For this reason, the present study examines the role of complexity-based descriptors for identifying stochastic perturbations in time series and analyzes how these metrics respond to different noise regimes across heterogeneous dynamical systems. A supervised learning approach based on complexity descriptors was developed to analyze controlled perturbations in multiple signal types. Gaussian, pink, and low-frequency noise disturbances were injected at predefined intensity levels into the Rössler and Lorenz chaotic systems, the Hénon map, and synthetic electrocardiogram signals, while AR(1) processes were used for validation on inherently stochastic signals. From these systems, eighteen entropy-based, fractal, statistical, and singular value decomposition-based complexity metrics were extracted from either raw signals or reconstructed phase spaces. These features were used to perform three classification tasks that capture different aspects of noise characterization, including detecting the presence of noise, identifying the perturbation type, and discriminating between different noise intensities. In addition to predictive modeling, the study evaluates the complexity profiles and feature relevance of the metrics under varying perturbation regimes. The results show that no single complexity metric consistently discriminates noise regimes across all systems. Instead, system-specific relevance patterns emerge. Under given experimental constraints (data partitioning, machine learning algorithm, etc.), Approximate Entropy provides the strongest discrimination for the Lorenz system and the Hénon map, the Coefficient of Variation, Sample and Permutation Entropy dominate classification for ECG signals, and the Condition Number and Variance of first derivative together with Fisher Information are most informative for the Rössler system. Across all datasets, the proposed framework achieves an average accuracy of 99% for noise presence detection, 98.4% for noise type classification, and 98.5% for noise intensity classification. These findings demonstrate that complexity metrics capture structural and statistical signatures of stochastic perturbations across a diverse set of dynamic systems.

By systematically introducing small symmetry-violating terms into quantum gates, we examine whether such perturbations can reproduce the tomographic noise patterns observed on actual quantum hardware. This study fits into the broader context of quantum computing and quantum information processing, where realistic noise modeling remains essential for understanding device behavior.Our analytical framework mirrors Qiskit’s tomography procedures and shot model, using transparent noise control. Our results show that weak generator perturbations can reproduce average deviations in Pauli Transfer Matrices with high fidelity (given the recorded hardware noise characteristics). These findings suggest that symmetry-violating generators provide a phenomenological model for relevant classes of hardware noise. This approach offers a reproducible analytic baseline for testing noise hypotheses and calibrating simplified models against experimental data.Beyond this descriptive role, the appearance of stable symmetry-violating perturbations in the fitted models supports their physical relevance: they indicate that real devices operate slightly outside the symmetry-preserving regime assumed in standard noise models, and that these deviations leave identifiable signatures in tomographic reconstructions. Because the analytical framework isolates these violations at the generator level, it enables principled diagnosis of non-unitality and related effects. As a result, the model not only reproduces experimental process tomography data but also identifies which symmetry assumptions fail and by what magnitude, providing actionable information for calibration, simulation, and the development of symmetry-sensitive mitigation and error-correction strategies.