A Reinforcement Learning approach to discriminate unsafe devices in aggregate computing systems

Reinforcement learning is a machine learning approach that has been studied for many years, but particularly nowadays the interest about this topic has exponentially grown. Its purpose is to create autonomous agents able to sense and act in their environment. They should learn to choose optimal actions to achieve their goals, in order to maximise a cumulative reward. Aggregate programming is a paradigm that supports the large-scale programming of adaptive systems by focusing on the behaviour of the cluster instead of the singles. One promising aggregate programming approach is based on the field calculus, that allows the definition of aggregate programs by the functional composition of computational fields. A topic of interest related to Aggregate Computing is computer security. Aggregate Computing systems are, in fact, vulnerable to security threats due to their distributed nature, situatedness and openness, which can make participant nodes leave and join the computation at any time. A solution that enables to combine reinforcement learning, aggregate computing and security, would be an interesting and innovative approach, especially because there are no experiments so far that include this combination. The goal of this thesis is to implement a Scala library for reinforcement learning, which must be easily integrated with the aggregate computing context. Starting from an existing work, on trust computation in aggregate applications, we want to train a network, via reinforcement learning, which through the calculation of the gradient -- a fundamental pattern of collective coordination -- is able to identify and discriminate compromised nodes. The dissertation work focused on: 1. development of a generic Scala library that implements the reinforcement approach, in accord to an aggregate computing model; 2. development of a reinforcement learning based solution; 2. integration of the solution that allows us to calculate the trust gradient.