As deep learning models have become increasingly complex, it is critical to understand their decision-making, particularly in safety-relevant applications. In order to support a quantitative interpretation of an autonomous agent trained through Deep Reinforcement Learning (DRL) in the highway-env simulation environment, we propose a framework featuring three types of views for analyzing data: (i) episode timeline, (ii) frame by frame, and (iii) aggregated statistical analysis, also including heatmaps for a better spatial understanding. Our methodology allowed a novel, consistent description of the behavior of the agent. The main motivator for the taken action is typically the longitudinal distance from the second-closest and, to a lower extent, third-closest vehicle. In the overtakes, also the agent’s position in lanes becomes relevant. The analysis identified interesting patterns and an issue in the last frames of an episode, when the agent is unable to overtake the last two vehicles, arguably because of the lack of reference vehicles ahead. We observed a clear differentiation between attention and SHAP values (estimating the importance of each feature for each decision), reflecting the architecture of the neural network, where the first layer implements the attention mechanism, while the deeper ones make the actual decision. Attention focuses on the proximity of the ego, while the decision is taken on a wider horizon, denoting a valuable anticipation capability. To support research, the proposed framework is released as open source.

Binarization is an extreme quantization technique that is attracting research in the Internet of Things (IoT) field, as it radically reduces the memory footprint of deep neural networks without a correspondingly significant accuracy drop. To support the effective deployment of Binarized Neural Networks (BNNs), we propose CBin-NN, a library of layer operators that allows the building of simple yet flexible convolutional neural networks (CNNs) with binary weights and activations. CBin-NN is platform-independent and is thus portable to virtually any software-programmable device. Experimental analysis on the CIFAR-10 dataset shows that our library, compared to a set of state-of-the-art inference engines, speeds up inference by 3.6 times and reduces the memory required to store model weights and activations by 7.5 times and 28 times, respectively, at the cost of slightly lower accuracy (2.5%). An ablation study stresses the importance of a Quantized Input Quantized Kernel Convolution layer to improve accuracy and reduce latency at the cost of a slight increase in model size.

As timely information about a project’s state is key for management, we developed a data toolchain to support the monitoring of a project’s progress. By extending the Measurify framework, which is dedicated to efficiently building measurement-rich applications on MongoDB, we were able to make the process of setting up the reporting tool just a matter of editing a couple of .json configuration files that specify the names and data format of the project’s progress/performance indicators. Since the quantity of data to be provided at each reporting period is potentially overwhelming, some level of automation in the extraction of the indicator values is essential. To this end, it is important to make sure that most, if not all, of the quantities to be reported can be automatically extracted from the experiment data files actually used in the project. The originating use case for the toolchain is a collaborative research project on driving automation. As data representing the project’s state, 330+ numerical indicators were identified. According to the project’s pre-test experience, the tool is effective in supporting the preparation of periodic progress reports that extensively exploit the actual project data (i.e., obtained from the sensors—real or virtual—deployed for the project). While the presented use case concerns the automotive industry, we have taken care that the design choices (particularly, the definition of the resources exposed by the Application Programming Interfaces, APIs) abstract the requirements, with an aim to guarantee effectiveness in virtually any application context.

We propose a new, hierarchical architecture for behavioral planning of vehicle models usable as realistic non-player vehicles in serious games related to traffic and driving. These agents, trained with deep reinforcement learning (DRL), decide their motion by taking high-level decisions, such as “keep lane”, “overtake” and “go to rightmost lane”. This is similar to a driver’s high-level reasoning and takes into account the availability of advanced driving assistance systems (ADAS) in current vehicles. Compared to a low-level decision making system, our model performs better both in terms of safety and speed. As a significant advantage, the proposed approach allows to reduce the number of training steps by more than one order of magnitude. This makes the development of new models much more efficient, which is key for implementing vehicles featuring different driving styles. We also demonstrate that, by simply tweaking the reinforcement learning (RL) reward function, it is possible to train agents characterized by different driving behaviors. We also employed the continual learning technique, starting the training procedure of a more specialized agent from a base model. This allowed significantly to reduce the number of training steps while keeping similar vehicular performance figures. However, the characteristics of the specialized agents are deeply influenced by the characteristics of the baseline agent.

Datasets are key to developing new machine learning-based applications but are very costly to prepare, which hinders research and development in the field. We propose an edge-to-cloud end-to-end system architecture optimized for sport activity recognition dataset collection and application deployment. Tests in authentic contexts of use in four different sports have revealed the system’s ability to effectively collect machine learning-usable data, with an energy consumption compatible with the timeframe of most of the sport types. The proposed architecture relies on a key feature of the Measurify internet of things framework for the management of measurement data (i.e., .csv dataset management) and supports a workflow designed for efficient data labeling of signal timeseries. The architecture is independent of any specific sport, and a new dataset generation application can be set up in a few days, even by novice developers. With a view to concretely supporting the R&D community, our work is released open-source.