Formal Synthesis of Collaborative Protocols for Joint Learning and Control
Technical point of contact: Chris Orlowski, DARPA (TTO)
Period of activity: 2015-2018
Overview of the Project
The objective of this project is to develop methods and tools for the formal specification and automated synthesis of protocols for collaborative control and learning for provably trustworthy operation of teams of autonomous systems in dynamic, uncertain and a priori unknown environments.
Our recent work developed on a mathematically sound formalism for joint control and learning in unknown stochastic environments. This projects we extends this joint learning and temporal logic-constrained control formalism to environments with nondeterministic as well as stochastic components. The algorithm development is in two branches: We generalize the existing algorithms to a two-player, stochastic game setup. (ii) We pursue correctness in joint learning and control by constraining run-time adaptation within a priori constructed permissive strategies that characterize allowable operating envelopes. Building on these results we aim to develop algorithmic support for distributed and collaborative joint learning and reactive control. On one side, we are adopting contract-based reasoning in order to obtain joint learning and temporal logic-constrained control tasks distributed over a network of agents. On the other, we are developing compositional joint learning and control protocol synthesis algorithms for Markov decision processes via distributed optimization.