The internet of Things (IoT) is attracting an increasing attention from both industry and academia, and will be a core component of the next generation 6G systems. In an IoT network, a potentially massive number of low-cost, low-complexity devices generate traffic to be reported to a common receiver monitoring the situation. Relevant examples include asset tracking, industrial and environmental monitoring, smart cities, as well as cyber-physical systems. In such a scenario, the possibility to let a large amount of concurrent transmissions coexist is key to improve the spectral efficiency and reliability of the next generation communication systems. Small data units sporadically transmitted by a vast population of terminals with low- to very-low duty cycle, poses unprecedented challenges to the multiple access design. Coordinating the transmission entails an overhead that may grow unbearably with the total population size and not with the number of active users, which might be order of magnitude lower. Further, such terminals might be battery powered and hence low energy consumption is key.

The aim of the thesis is to evaluate how machine learning can aid and possibly enhance the detection of small data packets transmitted asynchronously and uncoordinatedly. In modern random-access solutions, several overlapping packets may anyhow be decoded, thanks to powerful forward error correction, interference cancellation and other advanced signal processing techniques. In these scenarios, packet detectors are required to operate in a very harsh environment with unfavorable signal-to-noise and interference ratio. Additionally, detection algorithms that are theoretically optimal are difficult (if not impossible) to devise, due to the exponential number of possible interference cases that may arise in an uncoordinated and asynchronous access policy. Machine-learning algorithms appear to be very suited to such cases and in the thesis we would like to explore what relevant supervised learning algorithms may be beneficial, taking into consideration relevant trade-offs, in terms of e.g. complexity, detection performance, etc.