Type 1 diabetes is a growing global health challenge. Standard clinical practice often relies on manual insulin injections, which can lead to suboptimal glucose regulation. Recent advancements have shifted focus towards Artificial Pancreas systems, integrating continuous glucose monitoring with automated insulin delivery. This work presents a quantitative comparison of four Deep Reinforcement Learning algorithms for autonomous glycemic regulation via insulin injection: DDPG, PPO, SAC, and TD3. The validation is conducted using the Hovorka model, in presence of uncertainties on number, time and amount of meals. Results show that all four controllers are able to maintain blood glucose levels within the target range. The TD3 algorithm outperforms the others in terms of several key performance indicators such as time in range, time in hypo/hyperglycemia and total insulin usage, while also exhibiting fewer hyperglycemic episodes compared to prior works in academic literature.

Epilepsy is a non-communicable neurological disorder that causes recurrent and unprovoked seizure. Ideally, predicting seizures would represent a step forward in making life easier for those who suffer from epilepsy. This work aims to predict the occurence of epileptic seizures via a novel approach that combines Dynamic Mode Decomposition (DMD), that is a data-driven modelling technique for dynamical systems with a deep learning classifier, that is based on a convolutional neural network. The resulting two-stage data-driven predictor was tailored for the analysis of encephalographic (EEG) data. The validity analysis of the approach is carried out over the CHB-MIT Scalp EEG Database, demonstrating its applicability for seizures’ recognition and prediction on real clinical data.

Accurate path tracking is one of the main challenges autonomous vehicles have to deal with. It is known that, when dealing with real hardware systems, the presence of parametric uncertainty and unmodelled aspects of the system dynamics affects all model-based control approaches, hindering their nominal performance guarantees. To compensate for this issue, data-driven schemes have drawn significant attention from the scientific community thanks to their inherent ability to learn from experience, thus automatically compensate for system uncertainties and time-varying behaviours. This work aims to develop a reinforcement learning-based longitudinal and lateral dynamics control introducing mismatch and motion penalization metrics, validating the resulting controller in a simulated Formula SAE skid-test scenario employing the Sapienza Fast Charge Formula SAE Electric Racing Team dynamical model.

Autonomous driving has emerged as a technology to revolutionize the future of transportation and completely re-define the landscape of road systems. Accurate control algorithms are crucial to ensure the safety and efficiency of autonomous vehicles; in particular, incorporating actuator dynamics into the vehicle dynamics model can improve the response of the system to control commands, with clear safety implications. This work proposes pulse width modulation-based control strategies for the steering and brake actuators of a Formula SAE driverless vehicle. Extensive simulation tests and real-world experiments on a Formula SAE skid-test scenario validate the proposed approach.

In the context of Mobile Augmented Reality, satisfying the challenging users’ requirements about Quality of Service and Quality of Experience is not an easy task due to the limited computing capabilities of mobile devices, and the rapid, free movement of users within the environment. To deal with these issues, graphical computations are typically offloaded from mobile devices to edge servers. While traditional offloading strategies rely on static optimization or heuristics, this work proposes a multi–input data–driven dynamic control of uplink power and image compression rate, introducing a Policy Broadcasting Deep Reinforcement Learning approach, based on the Deep Deterministic Policy Gradient algorithm. The proposed solution is aimed at matching the challenging Quality of Service constraints, in terms of maximum round–trip latency and minimum resolution accuracy, while minimizing the energy consumption. Simulations show the effectiveness and scalability of the proposed approach for real–time applications.

Model Predictive Control (MPC) is a popular control strategy that relies on the availability of a prediction model to estimate future system trajectories over a finite time horizon. Recently, researchers have introduced Neural Networks (NNs) into the MPC framework for the development of data-driven prediction models. In MPC, the control actions are computed by solving iteratively, at each time-step, an optimization problem subject to state and input constraints. Finding the optimal solution to such a problem is a crucial challenge in the data-driven setting, due to the complexity and black-box nature of data-driven models such as NNs. This paper addresses this challenge by proposing a hierarchical deep NN formed by a set of cascading one-step NN predictors whose combination constitutes an interpretable prediction model over the entire prediction horizon. Thanks to the proposed NN architecture, it is shown that the resulting optimal control problem is tractable, as it can be solved by employing efficient iterative algorithms, and interpretable, so that input and state constraints can be enforced seamlessly. The effectiveness of the proposed method is validated through numerical simulations. Note to Practitioners—Model Predictive Control (MPC) is a widely used methodology in the industry which typically relies on the availability of a model in the form of step response, transfer function or state-space models. In some cases, the explicit model might not be available or its accuracy may be not sufficient for the required closed-loop performance. This paper aims to develop a simple and practical framework for deploying a model-free data-driven MPC solution based on deep learning. This objective is pursued by suggesting a novel approach using simple neural networks in a cascading interpretable structure. Such networks are used to predict the one-step evolution of the system, and their cascade represents the MPC prediction model over an arbitrary long prediction horizon. We characterize such a neural model focusing on its interpretability and tractability, deriving the resulting optimal control problem to be solved in a receding horizon strategy. We then show that the MPC optimization can be solved efficiently using highly efficient iterative algorithms that can be implemented in practice. Numerical simulations involving the use of the Alternating Direction Method of Multipliers (ADMM) algorithm show its effectiveness for both linear and nonlinear systems.

Topology optimization in decentralised federated learning settings enables the design of policies aimed at minimizing the number of communication rounds needed to reach algorithmic convergence. Given a federation of autonomous agents, finding the optimal topology which guarantees that the underlying graph is connected is still an open issue. This paper proposes a novel energy-aware topology optimization algorithm with the goal to derive an optimal topology which maximizes the algebraic connectivity of the corresponding graph in presence of energy and communication constraints. The effectiveness of the proposed approach is validated in the context of a consensus-based federated learning algorithm over an e-Health scenario.