Smart Self-Adaptive Cyber-Physical Systems: How can Exploration and Learning Improve Performance in a Partially Observable Multi-Agent Context?

Ana Petrovska, Malte Neuss, Sebastian Bergemann, Martin Büchner und Ansab Shohab

ADAPTIVE 2021: The Thirteenth International Conference on Adaptive and Self-Adaptive Systems and Applications,

April 2021


Cyber-physical systems (CPSs) are software-intensive systems that are embedded in the physical world to monitor, control and coordinate a variety of processes in both the physical and the digital world. As a result, they often operate in complex, dynamic, and unanticipated environments with various potential sources of run-time changes and uncertainties, that could potentially lead the CPSs to faults, and even to complete system failures. To cope with these changes, the systems should have the capabilities to self-adapt in order to continue meeting their functional specifications. In this paper, we investigate how creating self-adaptive CPSs which are able to collaborate and learn in a dynamic, partially observable, multi-agent context, can not only preserve but also improve the performance, despite all the changes introduced to the system at run-time. We evaluate the proposed methodology on an in-house developed, multi-agent system from the robotics domain.

Stichworte: self-adaptive systems, cyber-physical systems, collaboration, learning, partial observability, Model-based Systems Engineering, MbSE