Junxing Yang

Type

Text

Type

Dissertation

Advisor

Grosu, Radu | Smolka, Scott A. | Stoller, Scott D. | Tiwari, Ashish.

Date

2017-08-01

Keywords

Computer science

Department

Department of Computer Science.

Language

en_US

Source

This work is sponsored by the Stony Brook University Graduate School in compliance with the requirements for completion of degree

Identifier

http://hdl.handle.net/11401/78214

Publisher

The Graduate School, Stony Brook University: Stony Brook, NY.

Format

application/pdf

Abstract

Collective Autonomous Systems (CAS) is a class of systems constituted by multiple autonomous agents that cooperate on completing mission goals while satisfying certain safety constraints. We present Optimal Safe Planning and Control (OSPaC), a framework for optimal planning and control that features the use of a uniform cost function for formulating mission goals and safety constraints. We first introduce ARES, an algorithm for generating optimal plans that take a system to a goal state having a desired cost. ARES chooses the prediction horizon dynamically based on the progress toward the cost goal. Based on ARES, we also present Adaptive-Horizon MPC (AMPC), a new and effective formulation of model predictive control. We assess the effectiveness of ARES and AMPC using the case study of V-formation the goal of which is to bring a flock of birds (UAVs) from a random initial state to a state exhibiting a V-formation, with collision freedom as the safety constraint. Our experiments show that ARES and AMPC succeed with high probability. Inspired by the emerging problem of CPS security, we also introduce controller-attacker games, a two-player stochastic game where the two players, a controller and an attacker, have antagonistic objectives. In V-formation games, a particular form of controller-attacker games, the attacker attempts to prevent the controller from maneuvering a flock into a V-formation. Our experiments show that an AMPC controller adapts to the severity of attacks by choosing the prediction horizon suitably, and our results show that an intelligent attacker can significantly outperforms its naive counterpart. We also present DAMPC, a distributed adaptive-neighborhood AMPC algorithm. In DAMPC, agents run AMPC with local information to find actions for all their neighbors. A communication round is performed by local message exchange to reach agreement on a global solution. DAMPC dynamically adjusts the neighborhood size based on the current state-cost. This improves efficiency without compromising convergence. Finally, we present the integration of planning and control, where ARES generates optimal plans offline and AMPC follows the plans online. The integration improves the performance of AMPC with a computationally cheaper cost function and optimization algorithm, while still benefiting from its robustness to environment disturbances. | 95 pages

Download

Share

COinS