In the past few years, Federated Learning (FL) has emerged as an effective approach for training Neural Networks (NNs) over a computing network while preserving data privacy. Most existing FL approaches require defining a priori (1) a predefined structure for all the NNs running on the clients and (2) an explicit aggregation procedure. These can be limiting factors in cases where predefining such algorithmic details is difficult. Recently, NEvoFed was proposed, an FL method that leverages Neuroevolution running on the clients, in which the NN structures are heterogeneous and the aggregation is implicitly accomplished on the client side. Here, we propose MFC-NEvoFed, a novel approach to FL that does not require learning models, i.e., neural network parameters, to be distributed over the networks, thus taking a step toward security improvement. The only information exchanged in client/server communication is the performance of each model on local data, allowing the emergence of optimal NN architectures without needing any kind of model aggregation. Another appealing feature of our framework is that it can be used with any Machine Learning algorithm provided that, during the learning phase, the model updates do not depend on the input data. To assess the validity of MFC-NEvoFed, we test it on four datasets, showing that very compact NNs can be obtained without drops in performance compared to canonical FL. Finally, such compact structures allow for a step toward explainability, which is highly desirable in domains such as digital health, from which the tested datasets come.

Continuous control tasks often involve high-dimensional, dynamic, and non-linear environments. State-of-the-art performance in these tasks is achieved through complex closed-box policies that are effective, but suffer from an inherent opacity. Interpretable policies, while generally underperforming compared to their closed-box counterparts, advantageously facilitate transparent decision-making within automated systems. Hence, their usage is often essential for diagnosing and mitigating errors, supporting ethical and legal accountability, and fostering trust among stakeholders. In this paper, we propose SMoSE, a novel method to train sparsely activated interpretable controllers, based on a top-1 Mixture-of-Experts architecture. SMoSE combines a set of interpretable decision-makers, trained to be experts in different basic skills, and an interpretable router that assigns tasks among the experts. The training is carried out via state-of-the-art Reinforcement Learning algorithms, exploiting load-balancing techniques to ensure fair expert usage. We then distill decision trees from the weights of the router, significantly improving the ease of interpretation. We evaluate SMoSE on six benchmark environments from MuJoCo: our method outperforms recent interpretable baselines and narrows the gap with non-interpretable state-of-the-art algorithms.

Prompt engineering has proven to be a crucial step in leveraging pretrained large language models (LLMs) in solving various real-world tasks. Numerous solutions have been proposed that seek to automate prompt engineering by using the model itself to edit prompts. However, the majority of state-of-the-art approaches are evaluated on tasks that require minimal prompt templates and on very large and highly capable LLMs. In contrast, solving complex tasks that require detailed information to be included in the prompt increases the amount of text that needs to be optimised. Furthermore, smaller models have been shown to be more sensitive to prompt design. To address these challenges, we propose an evolutionary search approach to automated discrete prompt optimisation consisting of two phases. In the first phase, grammar-guided genetic programming is invoked to synthesise prompt-creating programmes by searching the space of programmes populated by function compositions of syntactic, dictionary-based and LLM-based prompt-editing functions. In the second phase, local search is applied to explore the neighbourhoods of best-performing programmes in an attempt to further fine-tune their performance. Our approach outperforms three state-of-the-art prompt optimisation approaches, PromptWizard, OPRO, and RL-Prompt, on three relatively small general-purpose LLMs in four domain-specific challenging tasks. We also illustrate several examples where these benchmark methods suffer relatively severe performance degradation, while our approach improves performance in almost all task-model combinations, only incurring minimal degradation when it does not.

Modern agriculture has enabled food production for nearly eight billion people, yet plant diseases and climate change continue to threaten food security. Beans, a key nutritional crop worldwide, are vulnerable to diseases such as bean rust and angular leaf spot, which significantly reduce yield. Early detection is critical for effective treatment. While existing methods leverage cloud-based Deep Learning (DL) models for plant disease classification, they often require substantial computational resources. In this work, we propose an edge-based solution by deploying a quantized MobileNetV2 model on a resourceconstrained embedded device. We compare the performance of the full-precision (Float 32-bit) model and its quantized (Int 8-bit) counterpart deployed on a microcontroller unit (MCU), targeting Internet of Things (IoT) scenarios. Our study analyzes model accuracy, inference speed, and memory utilization, focusing on peak RAM and flash memory requirements. The results show that quantization significantly reduces memory footprint and inference time while maintaining competitive classification accuracy. These findings highlight the potential of quantization to enable efficient, sustainable, and deployable TinyML models for plant disease detection at the edge.

Abstract In the current Artificial Intelligence (AI) landscape, addressing explainability and interpretability in Machine Learning (ML) is of critical importance. In fact, the vast majority of works on AI focus on Deep Neural Networks (DNNs), which are not interpretable, as they are extremely hard to inspect and understand for humans. This is a crucial disadvantage of these methods, which hinders their trustability in high-stakes scenarios. On the other hand, interpretable models are considerably easier to inspect, which allows humans to test them exhaustively, and thus trust them. While the fields of eXplainable Artificial Intelligence (XAI) and Interpretable Artificial Intelligence (IAI) are progressing in supervised settings, the field of Interpretable Reinforcement Learning (IRL) is falling behind. Several approaches leveraging Decision Trees (DTs) for IRL have been proposed in recent years. However, all of them use goal-directed optimization methods, which may have limited exploration capabilities. In this work, we extend a previous study on the applicability of Quality–Diversity (QD) algorithms to the optimization of DTs for IRL. We test the methods on two well-known Reinforcement Learning (RL) benchmark tasks from OpenAI Gym, comparing their results in terms of score and “illumination” patterns. We show that using QD algorithms is an effective way to explore the search space of IRL models. Moreover, we find that, in the context of DTs for IRL, QD approaches based on MAP-Elites (ME) and its variant Covariance Matrix Adaptation MAP-Elites (CMA-ME) can significantly improve convergence speed over the goal-directed approaches.

In the past few years, Federated Learning (FL) has emerged as an effective approach for training neural networks (NNs) over a computing network while preserving data privacy. Most of the existing FL approaches require the user to define a priori the same structure for all the NNs running on the clients, along with an explicit aggregation procedure. This can be a limiting factor in cases where pre-defining such algorithmic details is difficult. To overcome these issues, we propose a novel approach to FL, which leverages Neuroevolution running on the clients. This implies that the NN structures may be different across clients, hence providing better adaptation to the local data. Furthermore, in our approach, the aggregation is implicitly accomplished on the client side by exploiting the information about the models used on the other clients, thus allowing the emergence of optimal NN architectures without needing an explicit aggregation. We test our approach on three datasets, showing that very compact NNs can be obtained without significant drops in performance compared to canonical FL. Moreover, we show that such compact structures allow for a step towards explainability, which is highly desirable in domains such as digital health, from which the tested datasets come.

Batteryless edge devices are extremely resource-constrained compared to traditional mobile platforms. Existing tiny deep neural network (DNN) inference solutions are problematic due to their slow and resource-intensive nature, rendering them unsuitable for batteryless edge devices. To address this problem, we propose a new approach to embedded intelligence, called Fast-Inf, which achieves extremely lightweight computation and minimal latency. Fast-Inf uses binary tree-based neural networks that are ultra-fast and energy-efficient due to their logarithmic time complexity. Additionally, Fast-Inf models can skip the leaf nodes when necessary, further minimizing latency without requiring any modifications to the model or retraining. Moreover, Fast-Inf models have significantly lower backup and runtime memory overhead. Our experiments on an MSP430FR5994 platform showed that Fast-Inf can achieve ultra-fast and energy-efficient inference (up to 700x speedup and reduced energy) compared to a conventional DNN.

The evolution of symbolic communication is a longstanding open research question in biology. While some theories suggest that it originated from sub-symbolic communication (i.e., iconic or indexical), little experimental evidence exists on how organisms can actually evolve to define a shared set of symbols with unique interpretable meaning, thus being capable of encoding and decoding discrete information. Here, we use a simple synthetic model composed of sender and receiver agents controlled by Continuous-Time Recurrent Neural Networks, which are optimized by means of neuro-evolution. We characterize signal decoding as either regression or classification, with limited and unlimited signal amplitude. First, we show how this choice affects the complexity of the evolutionary search, and leads to different levels of generalization. We then assess the effect of noise, and test the evolved signaling system in a referential game. In various settings, we observe agents evolving to share a dictionary of symbols, with each symbol spontaneously associated to a 1-D unique signal. Finally, we analyze the constellation of signals associated to the evolved signaling systems and note that in most cases these resemble a Pulse Amplitude Modulation system.

Routing plays a fundamental role in network applications, but it is especially challenging in Delay Tolerant Networks (DTNs). These are a kind of mobile ad hoc networks made of, e.g., (possibly, unmanned) vehicles and humans where, despite a lack of continuous connectivity, data must be transmitted while the network conditions change due to the nodes’ mobility. In these contexts, routing is NP-hard and is usually solved by heuristic “store and forward” replication-based approaches, where multiple copies of the same message are moved and stored across nodes in the hope that at least one will reach its destination. Still, the existing routing protocols produce relatively low delivery probabilities. Here, we genetically improve two routing protocols widely adopted in DTNs, namely, Epidemic and PRoPHET, in the attempt to optimize their delivery probability. First, we dissect them into their fundamental components, i.e., functionalities such as checking if a node can transfer data, or sending messages to all connections. Then, we apply Genetic Improvement (GI) to manipulate these components as terminal nodes of evolving trees. We apply this methodology, in silico, to six test cases of urban networks made of hundreds of nodes and find that GI produces consistent gains in delivery probability in four cases. We then verify if this improvement entails a worsening of other relevant network metrics, such as latency and buffer time. Finally, we compare the logics of the best evolved protocols with those of the baseline protocols, and we discuss the generalizability of the results across test cases.