ActionWebs was a large CPS project active during 2009-2016, focussed on the foundational analysis and design of systems of actionable sensors.  These are systems that can be automatically controlled to gather information about their environment, yet their main role may involve a separate task potentially at odds with sensing.  Aircraft in an air traffic control (ATC) system, for example, are typically a key source of information about weather and flying conditions for ATC.  Another example is in automated buildings, in which sensors can be tasked with gathering information pertinent to localized control.

Motivated by the ITR project CHESS (Center for Hybrid and Embedded Systems and Software) and the DARPA SEC (Software Enabled Control) project, ActionWebs spanned the domains of model identification and control, as well as the design of the embedded software.  The topic of real time treatment of data in the control loop emerged as a key topic in the project. 

ActionWebs began with four research areas:  observe and infer with a viewpoint to planning and modifying action, distributed control, programming the ensemble, and closing the loop around sensor networks.  As the project developed, these were refined into six key research themes, which along with the two driving applications of Next Generation Air Transportation, and Energy-Efficient, High Productivity Buildings, became the core of the project.  Key Outcomes in each of these areas are described in the following.

(1) Learning and control.  In ActionWebs we addressed the problems of joint inference and control head-on, with key results in the design of control algorithms which blend machine learning with formal methods from verification: the high performance achievable using statistical regression benefits from the traditional stability and reachability proofs available from control theory.  Results included Safe Learning, in which machine learning algorithms are used to increase performance inside reachable sets.

(2) Stochastic hybrid systems:  In many practical application scenarios featuring cyber-physical systems, the information available for control is often an imperfect representation of the true system state.  This can arise from limited sensor placements, decentralized information patterns, communication drop-outs, as well as environment noise.  Using partially observable stochastic hybrid systems as an abstraction for such scenarios, we designed algorithms and computational methods for probabilistic safety and reach-avoid problems formulated as partial information stochastic optimal control problems with multiplicative or sum-multiplicative payoffs.

(3) Distributed and decentralized game theory.  In ActionWebs, we have developed methods for incentive design in multiple player games, within the frameworks of mean field games, principal-agent problems, dynamic mechanism design, and linear quadratic games. 

(4) Communications and control. We are designing novel control protocols for multiple vehicle interaction in a NextGen ATC setting, under realistic models of communication links between vehicles.  We built on a deep understanding of the interactions between the control and network layers to ensure the stability and efficiency of critical networked control systems.

(5) Data representation, visualization, and access for real time control.  ActionWebs designed a sensing and information management system for building-scale sensor networks, as a basis for real time control.  Our research has focussed on broad spectrum monitoring of the distributed physical infrastructure in sMAP, data-driven modeling and control, and human-centered building applications.

(6) Timing analysis for real time embedded systems.  The systems of interest in ActionWebs differ from passive sensor networks in their aim to control the physical systems in which they are embedded. Thus, they tightly integrate physical dynamics with computation and networking, and hence require abstractions for computation and networking that integrate temporal dynamics and concurrency.   ActionWebs supported the development of a programming model called Ptides for the model-based design of event-triggered real-time distributed systems.  Ptides is based on a discrete-event formalism that provides the necessary model for time and concurrency.

ActionWebs research has been applied mainly in two application areas:

Air transportation research: Secure and fault tolerant ATC protocols were developed which take into account the communication delays in information transmission.  Algorithms which use ATC data in real time to identify and characterize network delay modes, aircraft engine performance, and weather penetration by aircraft, and associated control schemes were developed.  Algorithms for integrated aircraft arrival and departure control were designed and implemented in Boston Logan. 

Energy efficient buildings research: ActionWebs helped to formalize a model-based control approach within buildings, which has been a predominantly 'set point based control' paradigm in the past.  ActionWebs designed Learning Based Model Predictive Control (LBMPC)  as a foundation for bringing modern control techniques into energy efficient buildings, and built data management tools for buildings such as sMAP and XBOS.  The Berkeley Retrofitted and Inexpensive HVAC Testbed for Energy Efficiency (BRITE) was developed, initially across two buildings on the Berkeley campus and now extended to several of the campus buildings as living labs for the energy efficiency and high productivity buildings. 

Last Modified: 01/14/2017
Modified by: Claire J Tomlin