Keeping Intelligence under Control

Modern software systems, such as smart systems, are based on a continuous interaction with the dynamic and partially unknown environment in which they are deployed. Classical development techniques, based on a complete description of how the system must behave in di erent environmental conditions, are no longer e ective. On the contrary, modern techniques should be able to pro- duce systems that autonomously learn how to behave in di erent environmental conditions. Machine learning techniques allow creating systems that learn how to execute a set of actions to achieve a desired goal. When a change occurs, the system can autonomously learn new policies and strategies for actions execution. This exibility comes at a cost: the developer has no longer full control on the system behaviour. Thus, there is no way to guarantee that the system will not violate important properties, such as safety-critical properties. To overcome this issue, we believe that machine learning techniques should be combined with suitable reasoning mechanisms aimed at assuring that the decisions taken by the machine learning algorithm do not violate safety-critical requirements. This paper proposes an approach that combines machine learning with run- time monitoring to detect violations of system invariants in the actions execution policies.