Probably Approximately Correct Protocols for Reactive Control and Learning

Jump to: navigation, search

Technical point of contact: Purush Iyer, ARO
Period of activity: 2015-2018

Overview of the Project

The objective of this project is to develop decision-making algorithms for autonomous and intelligent systems that jointly learn and react in environments with stochastic as well as adversarial uncertainties. The algorithms will be not only efficient in learning in terms of their use of samples, time, and space (i.e., in the traditional probably approximate correctness "PAC" sense) but also provably correct (by synthesis) with respect to rich temporal logic mission specifications. We balance theoretical development with immediate relevance by using shared decision-making between an human operator and autonomous mobile vehicles as a case study. caption

We partition the work into three research thrusts:

  • Probably approximate correctness in joint learning and temporal logic-constrained reactive synthesis: How can we adapt the notions of probably approximate correctness for safety-critical systems operating in environments with both stochastic uncertainties and adversarial opponents subject to temporal logic specifications?
  • Quantitative trade-offs, resilience and regret in joint learning and synthesis: How can we leverage the formalism from Thrust I to investigate exploration vs. exploration trade-offs, resilience of the joint protocols to changes between the design and deployment domains, and unconventional interactions between the controlled system and its environment and opponents?
  • Applications in shared control: We put the interplay between learning and reactive control into a concrete context and demonstrate the utility of our results through a case study in which a human operator works with unmanned systems with various levels of autonomy capabilities.