Large-Scale Optimization and Control

The cluster will develop basic theory and methodology for distributed optimization, learning
and decision-making in large scale dynamic systems. This is essential to efficiently and
reliably operate infrastructure networks for transportation, communications, data,
electricity, heat and water, as well as smart cities and health care. The main research
challenges are in the intersection between optimization, control, statistics, machine learning
and economics.

Vision for the Cluster

Background: Our modern society is critically dependent on large-scale networks for services such as
energy supply, transportation and communications. The design and operation of such networks is
becoming increasingly complex, due to their growing size, autonomy and heterogeneity. To address
these difficulties, a systematic theory and methodology for large-scale optimization and control is
needed. The cluster is addressing this challenge, building on leading academic expertise in related
areas, in combination with strategic industrial and international partnerships.

Objective: New theory and methodology will be developed in the form of algorithms and analysis
methods for dynamic operation of large scale systems. Optimization methods will be designed to
operate in real time interaction with the physical environment. For scalable solutions, this means
that data needs to processed in a distributed fashion, where local units for sensing, computation and
actuation are coordinated via physical infrastructure as well as communication networks and
broadcast technology. A relevant scenario is transmission and distribution of electrical power, where
high penetration of wind and solar power needs to be balanced by flexible demand response
triggered by economic incentives, with coupling to district heating networks and gas networks.
Another scenario is traffic control, where efficient methodology is needed to coordinate routing and
speed control for individual cars with conflict resolution at intersections and coordination of traffic
flow at a higher level.

Connection to other WASP Clusters: Many other WASP clusters are devoted to networks of autonomous
systems, such as vehicles and robots, designed to interact with other agents in a complex environment. Here
optimization is an important tool and scalability is a central issue. In particular, the cluster on
Perception and Learning relies heavily on scalable methods for optimization. Another example is the cloud
Autonomous Cloud and Networks which gives an implementation for methods developed in this cluster,
while theAutomated Transport Systems is focusing on an important application, where methods for
large scaleoptimization and control play an important role.

Future Demonstrations: This cluster is closely related to the goals stated for the WASP WARA-CAT
research arena. However, the truly large-scale aspects of the cluster will be hard to test in an
artificial experimental environment. As a consequence, opportunities for full scale demonstrations
will rely on collaborations with partners outside WASP, such as networks operators for traffic, power
or district heating.

Research Challenges

Scalable optimization in a dynamic environment

Optimization is already a widely used technology for dynamic control. The premier example is Model
Predictive Control based on centralized information, which is widely used, but far from scalable.
On the other hand, there is also a rapid growth of algorithms designed for scalability, but not in the
context of dynamical systems. The challenge is to combine the two, such that stability and dynamic
performance can be guaranteed, even when the algorithmic dynamics of scalable optimization are
strongly coupled with dynamics of physics, humans and communications.

Global versus local information

The theory for control and decision-making is relatively straightforward when full information about
the system is available. Another widely studied case is where information is limited, but localized and
fixed. This case is much more complicated to analyze, but good practical solutions often exist. Our
challenge is to develop scalable methods for control based on propagation of partial information,
with provable bounds on dynamic performance. In this setting, global broadcast information could be
combined with local information kept for internal use.

Adaptation and Learning

The rapid growth of machine learning algorithms during the past decades has drastically changed the
way we approach modeling and decision-making in autonomous systems. Most likely, this
development has only just started. However, most of the learning theory has been developed
without explicit regard for dynamics. For efficient integration of the new learning algorithms with
control of dynamical systems we will build on established theory for adaptive control and systems
identification, where stability and dynamic performance is explicitly taken into account.

Control through economic incentives

Many applications have a strong presence of independent decision-makers, responding to price
variations and individual preferences. The integration and coordination of such decision-makers will
be studied for a resource sharing network, where time-varying local demands are to be balanced by
time-varying supplies elsewhere in the network. The actions of each node are governed by a desire
to optimize a given utility function. In addition to local production and consumption, the exchange
between nodes is limited by capacity constraints of the network links. Our objective is to understand
and exploit the interaction between node demands and network dynamics. Privacy, security and
fairness will be essential factors in evaluation of achievable performance.

Industrial Challenges

There are numerous industrial challenges relating to this cluster. ABB is selling hardware and
software for automation of industrial processes, where optimization plays a key role. Industrial
trends are increasingly facilitating integration and data exchange between low level control loops
and high level process management and business decisions. This is increases the need for scalable
optimization and control. Cloud technology (for example provided by Ericsson) stimulates this
development further. Moreover, the “sharing economy” is creating new business models, where
information exchange and price mechanisms enable more efficient solutions to many societal needs.
For industry, this means new opportunities. At the same time, also the role of government is
becoming increasingly complex, creating a demand for analytical tools as provided in this cluster.

Sub-projects

Large-Scale Optimization for Distributed Control
Shervin Parvini Ahmadi, LiU

This PhD project will investigate network topology such as hierarchical network structure, e.g.
chordal graphs, for scalable computations with applications to control in a wide sense, e.g. system
identification, machine learning, control design, and estimation. We will also study how privacy affects
the achievable performance and what amount of communication that is needed between different
computational agents in order to achieve the most efficient overall computations. Different type of
computational platforms will be investigated. We will also develop efficient optimization code for
generic problem formulations. Relevant applications are within infrastructure networks for traffic,
water, gas, electricity, and building control. Industrial contacts have been established with
ABB Corporate Research in Switzerland.

Advanced real-time planning and decision making for autonomous systems
Kristoffer Bergman (industrial student), SAAB and LiU
This project aims at developing state-of-the-art real-time planning and decision making algorithms
that we believe will play a key-role of the “intelligence” of future autonomous systems. We would
like to, as far as possible, to work model based and minimize situation and platform specialization, as
well as operator involvement. The algorithms should themselves find solutions, i.e., a sequence of
decisions in time, to the given problems formulated in the form of a user-defined mission objective,
constraints on behaviors and actions, and a model of relevant parts of the world (updated in realtime
from observations and communication) where the platform acts.

Control using Distributed Information
Hamed Sadeghi, Lund University
Motivated by applications in infrastructure networks (mainly traffic and transportation) we are
studying how network flows can be optimized using distributed controllers. Existing results for linear
systems with an H-infinity objective will be generalized to accommodate non-linear flow constraints
and other convex objectives. Tradeoffs between congestion reduction and shortest path solutions
will be studied. During the initial phase, we will study how desirable static performance objectives
can be met using distributed feedback. Objectives and constraints appearing in traffic and
transportation networks will be emphasized. The next step is to design the dynamic properties of the
distributed controllers, while keeping the static properties intact.

Integrated verification, optimization and learning
Fredrik Hagebring, academic PhD, Chalmers
This project focuses on combining formal verification and optimization in correct-by design control
synthesis. The synthesis is based on specifications of the desired system behavior that may include
performance, safety and liveness properties. Optimality in terms of cost, performance and energy is
also crucial. To get a holistic view, correctness and optimality need to be further related. The goal is
to obtain a unified framework that guarantees optimal and correct behavior of autonomous systems,
including both controllable and uncontrollable/spontaneous dynamic behavior. Robustness
concerning uncertainties and adaptation/learning based on changes in decisions and system
behavior will be considered.

Towards the autonomous mine
Max Åstrand, industrial PhD, ABB and KTH
Underground mining is a cost-intensive process where the costs increase as production goes deeper.
To make production more efficient the mining industry turns to automation and autonomous mining
machinery. However, there are still many open problems that need to be addressed to realize the vision
of a fully autonomous mine. This project will contribute towards this vision starting by studying
automatic scheduling of an underground production fleet. Scheduling can be seen as the glue that
unites long-term goals with their autonomous realization, and is thus a crucial part of a fully autonomous mine.

Embedded optimization for real-time machine learning
Sarit Khirirat, academic PhD, KTH
An increasing number of our daily decisions are guided by computers that sense, infer and act
on the data that they observe. Many of these autonomous decision-making tasks are naturally posed as
optimization problems. Examples include finding the best parameters (“training”) of a deep neural network,
deriving an optimal decision rule for trading of volatile assets, or determining the best action of an autonomous
vehicle under dynamic constraints and uncertain observations. Real-time optimization is becoming a critical
technology for making better, faster and more informed autonomous decisions. In this project, we will develop
advanced optimization algorithms that are able to run in real-time on embedded hardware. We aim at novel
classes of optimization algorithms that can deliver optimal or near-optimal solutions with limited memory,
processing and energy resources. Particular attention will be given to algorithms that can exploit emerging
multi-core/multi-GPU embedded platforms. The project will blend applied mathematics for algorithm development,
implementation of these algorithms on emerging hardware architectures, and application of
real-time optimization to autonomous decision-making.