Uncertainty-Aware Coevolution Design of Self-Adaptive Cyber-Physical Systems

Self-adaptation in self-adaptive cyber-physical systems (self-CPSs) has gained great attention, mostly because of its envisioned benefits, that is, its flexibility and robustness in diverse known and unforeseen situations. Such benefits require a complex system design that must be able to handle two key challenges: 1) inevitable coevolution (reciprocally discovering their adaptation strategies and relationships and constantly coevolving with the environment) and 2) inherent internal and external (environment) uncertainty. Co-evolver addresses these two challenges, by exploring and exploiting the coevolution design of a self-CPS to a given level of maturity before deployment, and enabling the self-evolution of its coevolution strategy during operation. To achieve this, we draw on theories and technologies from model-based engineering, evolutionary computation, machine learning, and AI technologies. The goal is to ensure that a self-CPS is robust to known and unknown eventualities inherent in such systems. So far in the project, we have developed uncertainty quantification methods and machine learning techniques to improve the robustness of self-CPSs, with a particular focus on the autonomous driving domain. In addition, we have applied reinforcement learning and epigenetics evolutionary algorithms to build two new techniques for ensuring self-CPS robustness.

Self-adaptation in self-adaptive cyber-physical systems (self-CPSs) has gained great attention, mostly because of its envisioned benefits, that is, its flexibility and robustness in diverse known and unforeseen situations. Such benefits require a complex system design that must be able to handle two key challenges: 1) inevitable coevolution (reciprocally discovering their adaptation strategies and relationships and constantly coevolving with the environment) and 2) inherent internal and external (environment) uncertainty. Co-evolver addresses this challenge and will explore and exploit the coevolution design of a self-CPS to a given level of maturity before deployment and enable the self-evolution of its coevolution strategy during operation. To achieve this, we will draw on theories and technologies from model-based engineering, evolutionary computation, and machine learning. The goal is to ensure that a self-CPS is robust to the known and unknown eventualities inherent in such systems. The key scientific outcomes are 1) a multi-paradigm modelling framework (with a focus on coevolution and uncertainty modelling) for developing executable coevolution design models, 2) novel (co-)evolutionary algorithms and advanced applied studies on uncertainty-related theories and machine learning techniques to enable the continuous exploration and exploitation of coevolution designs, and 3) a comprehensive platform for evolving coevolution design models. The scientific impact of Co-evolver is multifaceted: advancing theoretical research in evolutionary computation, fostering a research niche in the coevolution design of complex systems, and leading the uncertainty modeling research community. Co-evolver will collaborate with NASA JPL&Caltech, USA (leading the development of a Computer Aided Engineering for Systems Architecture platform) and SOFTEAM Cadextan, France (leading the innovation and evolution of Modelio.org, a world leading open source modeling environment).