Background: In several complex systems, such as the transportation network, manufacturing or drug dose management, deriving an accurate physics-based model of the system is challenging due to the scale of the system and the uncertainties in the parameters and processes involved. Meanwhile, with the explosion of sensing and data processing capabilities, we can collect data from the systems at an unprecedented rate. This has resulted in a paradigm shift in control: from physics-based to data-driven. However, unless we prove safety of the data-driven approaches, we cannot employ them in safety-critical tasks such as autonomous driving, robot-human interactions or personalized medicine. Safe data-driven control problem: With the advent of data-driven control, the problem of safe data-driven control has gained an increasing attention. This problem can be formulated as control of a dynamical system under state and input constraints, with the caveat that the dynamical model or the constraints are a priori uncertain.Open challenges: Past work on safe data-driven control has mainly focused on the case of unknown dynamics but known constraints. The case in which the constraint sets themselves are being learned based on data is not sufficiently explored. This realistic case arises in several applications including autonomous driving, where the trajectories of nearby traffic are uncertain; autonomous robots operating in unexplored environments or along humans/other robots; or personalized medicine, where the effects of a drug on a given patient is a priori unknown. Goal and specific aims: The goal of this project is to develop the theory and algorithms for safe data-driven control, under uncertainties in the safe sets and their evolution. The specific aims are as follows:1. Developing fundamental understanding of safe data-driven control by characterizing its sample complexity;2. Modeling probability distributions capturing the uncertainties in an estimation module used to learn safe sets;3. Synthesizing controllers that ensure satisfaction of suitably defined notions of risk and safety;4. Verification of the algorithms on realistic simulation test-beds and on an in-house experimental platform. Expected outcome: I will contribute to both the scientific advancement of the problem and the training of researchers. I will publish the scientific results in strictly peer-reviewed conference and journal venues, across automatic control and machine learning communities, and publicize them through international talks. I will work closely with my trainees to build their fundamental optimization, control and learning technical skills and to support them in scientific communication. Additionally, I will organize workshops and online courses pertinent to the research objectives and outcomes, to exchange ideas and disseminate the research to a larger scientific community. Impact: The future automated systems, e.g. in intelligent transportation and personalized medicine, need to ensure safety and efficiency despite uncertainties arising from coupled subsystems, human interactions and partial data. Given the importance of this problem, the control and learning communities have been increasingly focusing on it, making breakthroughs on various formulations of this problem. My proposed research complements these efforts by bridging tools from stochastic control, online optimization and learning, all within my past contributions and expertise, to guarantee safety in a practically important but theoretically less-explored problem class, namely, uncertain safe sets. It advances fundamental understanding for this problem class, designs robust control algorithms leveraging this understanding, and offers a holistic approach by verifying and benchmarking the theoretical results on realistic test-beds.
Last updated:20.02.2023
