Correct-by-construction Control with Non-asymptotic Learning, Estimation and Detection in-the-Loop
Abstract
Many next generation navy missions require autonomous systems to operate robustly and predictably in highly dynamic environments. Such systems should be able to (i) react to cues/disturbances from the environment or to user commands regarding evolving mission objectives, (ii) adapt to changes or newly discovered facts in the environment, while also providing guarantees on safety and performance. A recent principled approach for designing autonomous systems is to employ correct-by-construction control synthesis techniques. These approaches start with a model of the system and a formal high-level specification describing the desired behavior, and by leveraging ideas from control theory (hybrid systems) and computer science (temporal logics, automata and game theory), automatically construct a controller that guarantees that the closed loop system satisfies the specification regarding safety and mission objectives given assumptions on the environment. When the assumptions on the system models and the environment are uncertain or misspecified, the resulting solutions can be overly conservative or even at the risk of being incorrect. The objective of this project is to develop the scientific foundation and associated algorithmic tools for the design of provably-correct autonomous systems that adapt themselves as the uncertainty is resolved via online learning, estimation and detection algorithms. In order to give guarantees on safety (as opposed to asymptotic properties like stability), we will focus on non-asymptotic techniques and their interplay with the control loop. If successful, the project will provide new tools and design methodologies that will enable fast development and deployment of trustworthy autonomous systems that can take full advantage of recent advances on online learning, estimation and detection. The key innovations of this research include: (1) Algorithms to compute the time to specification violation when specification violation is unavoidable due to unstructured disturbances or abrupt changes in the environment and algorithmic synthesis of controller maximizing this time. Such controllers will allow the estimation and learning modules to reduce the uncertainty in the newly encountered situation and also allow the resynthesis of new controllers that are guaranteed to operate in this new situation. When learning, estimation and detection algorithms with well-understood convergence rates are used, it will be possible to understand the trade-offs between the time required to learn and the time required to preserve safety. We will also consider controllers that maximize time to specification violation for different levels of relaxations of the specification. (2) Algorithms to learn and track slow changes in the environment or system dynamics and incremental synthesis techniques to modify and adapt existing controllers. This thrust is similar to adaptive control with the fundamental difference being the specifications include hard constraints. Therefore, classical asymptotic analysis in adaptive control with stability requirements is not applicable. We will develop theoretical foundations for non-asymptotic adaptation and correct-by-construction control by incorporating results from learning theory and partial information games. In particular with non-asymptotic learning the partial information gets completed with time and we plan to develop algorithms that exploits this structure for scalability. In a sense, our framework provides means for next generation autonomous systems to ???operate safe and adapt" with correctness guarantees as oppose to ???fail safe".
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 26, 2018
- Source ID
- N000141812501
Entities
People
- Necmiye Ozay
Organizations
- Board of Regents of the University of Michigan
- Office of Naval Research
- United States Navy