Distinguished Speech on Autonomous System by Turing Award Winner Joseph Sifakis
Date: 2019-05-16 Source:PCL
On May 15, Peng Cheng Lab welcomes Turing Award winner Prof. Joseph Sifakis to visit the Lab to deliver a distinguished speech entitled: “In Search of a Foundation for Next Generation Autonomous Systems”.
Prof. Sifakis received the Turing Award for his contribution to the theory and application of model checking in 2007. He is a member of the French Academy of Sciences, a member of the French National Academy of Engineering, a member of Academia Europea. He is also the founder of the Verimag laboratory in Grenoble. His current research interest is in system design, with formalization of autonomous system design as his focus.
Prof. Sifakis’ visionary speech is closely related to his current main research direction. In this distinguished talk, he introduced the path of realizing the next-generation autonomous system. Prof. Sifakis believes that the next generation of “autonomous systems” should be equipped with three important capabilities: 1) “broad intelligence”, which can replace human beings to deal with a series of dynamic and potentially conflicting goals; 2) the ability to interact with complex and unpredictable cyber physical environments; 3) the ability to coexist with the operators to provide safe and reliable application scenarios.
He argued that “autonomy” of the autonomous system should be embodied in the agent’s ability to make adaptive decisions and have self-awareness. The engineering foundation for implementing an autonomous system is to integrate model-based and data-based system design and verification methods. For the complexity of autonomous systems, Prof. Sifakis stated that there are still many challenges in the design of autonomous systems, including the perception, control, self-planning and target management of agents. From the perspective of structural design, the autonomous system needs to interact with the cyber physical environment. He believes that the research on autonomous systems will significantly broaden our horizons about artificial intelligence.
With regards to the concerns from the audience during the question-and-answer session about the difference between reinforcement learning and autonomous system, Prof. Sifakis believes that reinforcement learning is more about presenting a monolithic approach to specific problems and goals, while an autonomous system emphasizes solving complex problems within the system and managing potentially conflicting goals. In addition, because reinforcement learning is based on data, it is more about the machine’s “application without understanding” on specific issues. This is why such learning lacks of the mastery of the internal logic and rules of the system. The system based on this "black box mechanism" is relatively unstable. In contrast, the autonomous system requires the construction of the system theory and the overall stability and predictability of the system. Taking the automatic driving system as an example, the navigation issue could be solved by artificial intelligence. As to avoiding all potential obstacles and collisions, autonomous systems could provide more stable and secure system applications.