Be Prepared: Learning Environment Profiles for Proactive Rule-Based Production Planning

A key challenge in cyber-physical systems is to autonomously maintain system goals in the presence of uncertainties concerning the environment behaviour at runtime. Self-adaptivity has shown to be powerful to cope with this. It is usually based on a feedback loop that continuously evaluates the satisfaction of system goals and adapts the system in case of violations. However, most approaches rely on reactive adaptation, where the system triggers adaptations when goals are already violated. In contrast, proactive adaptation anticipates possible goal violations in the future and triggers adaptation in time. To enable this, the system maintains a model that allows for predicting the environment behaviour in the near future. However, these models are often heavyweight and resource-consuming to predict future behaviour (e.g. based on model checking). To overcome this problem, we present a lightweight approach for continuous learning of environment profiles and integrate it with resource-efficient rule-based adaptation. Moreover, we combine proactive adaptation that uses future predictions based on learned profiles, and reactive adaptation. We illustrate our ideas with an autonomous production system.