SYSTEM-OF-SYSTEMS THAT ACT
LOCALLY FOR OPTIMIZING
GLOBALLY
EU FP7 - SMALL/MEDIUM-SCALE FOCUSED RESEARCH PROJECT (STREP)
FP7-ICT-2013.3.4: ADVANCED COMPUTING, EMBEDDED AND CONTROL SYSTEMS
D) FROM ANALYZING TO CONTROLLING BEHAVIOUR OF SYSTEM OF SYSTEMS (SOS)
Cluster Meeting and Interim Review of FP7 Call 10 SoS projects
A Brief Overview of the Local4Global Project and progress made
in M1-M18
Elias Kosmatopoulos, CERTH
May 29, 2015, Florence, Italy
4
Local Global
Contact Information
For information regarding this Project: Check the Project Web-Site: http://local4global-fp7.eu
Participants
1 CERTH - Centre for Research and
Technology
2 ETHZ – Eidgenössische Technische
Hochschule Zürich
3 RWTH – RWTH Aachen University
4 IK4 – IK4 TEKNIKER
5 TUC – Technical University of Crete
6 TRV – TRANSVER GmbH
7 TUM – Technische Universtität
Muenchen
Project Acronym: Local4Global
Project Number: 611538
Project Start Date: October 2013
Duration: 3 Years
Funded by: EU FP7
Program Name:
EU FP7 - SMALL/MEDIUM-SCALE FOCUSED
RESEARCH PROJECT (STREP)
FP7-ICT-2013.3.4: ADVANCED COMPUTING,
EMBEDDED AND CONTROL SYSTEMS
D) FROM ANALYZING TO CONTROLLING
BEHAVIOUR OF SYSTEM OF SYSTEMS (SOS)
2
L4G: Plug-n-Play control for SoS
• Constituent systems (agents) operating within a SoS environment.
• Model-free control design :
• SoS dynamics can be extremely complex to model
• SoS dynamics may constantly changing
• Operate using local information (no need to deploy an “expensive”
infrastructure)
• Just-enough-learning/Just-enough-situation awareness
• deal with systems of constantly changing topology and structure
• detect and control evolutionary and emergent behavior
• operate locally in order to optimize globally
3
L4G: The Concept
• Is it possible to compute the agents’ actions that globally optimize the SoS
performance?
• Assumption 1 (non-realistic/non-SoS): agents “know” the dynamics of the
overall SoS.
• Assumption 2 (non-realistic/non-SoS): agents “know” what is happening
throughout the overall SoS.
• Optimal control and distributed optimization methodologies can provide
efficient solutions.
• Removal of Assumption 1: agents learn the SoS dynamics (not “all of them” but
just what they need to compute the optimal actions): Just-Enough-Learning
• Removal of Assumption 2: agents estimate what is happening throughout the
overall SoS (not “in each and every detail” but just what they need to compute
the optimal actions): Just-Enough-Situation Awareness
• Just-Enough-Learning and Just-Enough-Situation Awareness are possible only if
the control actions “sufficiently excite the SoS”: Control-for-Learning and
Learning-to-Control (C4L/L2C)
4
The elements of the Local4Global control approach
vs traditional control systems approach
• A self-learning mechanism to identify TSoS dynamics needed by each
constituent system, based on the least possible information from the whole
ensemble
⇛ system model
• A situation-awareness mechanism to estimate real TSoS state need by each
constituent system in order to calculate its own optimal actions, based on
the least possible information from the whole ensemble
⇛ state observer
• A distributed optimiser to determine the local control actions of each
constituent system, based on self-learning mechanism’s system identification
and self-awareness mechanism’s real state estimations
⇛ control strategy
• A Control-for-Learning and Learning-to-Control (C4L/L2C) mechanism to
refine the constituent systems’ local actions concurrently in an attempt to:
• optimise the TSoS performance at the global level and
• maximise the learning capabilities and situation awareness of the constituent
systems
⇛ upper hierarchical control layer
5
The Local4Global pursued response to TSoS
control vs traditional control approaches
6
Constituent system embedded software needs
The Local4Global software implementation approach
7
Local4Global: Use Cases
Goal: 30% improvements using real-life experiments
Cooperative Traffic SoS (Munich)
• Concurrently control traffic lights and
individual vehicle speeds
• SoS of constantly changing
structure/topology
• Only local information
• Emergent behavior
• Humans (drivers) in the loop
Energy Efficient Building SoS (Aachen)
• Concurrently control all energyinfluencing elements
• SoS of complex hierarchy/topology
• Only local information
• Emergent behavior
• Humans (occupants) in the loop
• Applicable to Every Day Buildings (with
no need to deploy an expensive
infrastructure) and Smart Grids
WP2: L4G-to-SoS Requirements
Work structure and outcomes
• Task 2.1
• Goal: Modelling and analysis requirements identification for the development of
Local4Global control theoretical tools suitable for generic TSoS
• Outcomes: Identification of TSoS modelling and analysis requirements; Development of a
systematic domain-independent work framework to assist TSoS requirements identification
and control; Development of a systematic domain-independent lexicon framework for the
representation of TSoS aspects that are essential from a control point of view
• Task 2.2
• Goal: Software/hardware requirements identification for the implementation of the
Local4Global control approach to generic TSoS
• Outcomes: Identification of functional and other requirements for implementing the
Local4Global control approach; Study and modelling of the interaction between embedded
devices and cloud framework; Identification of the interface between embedded devices and
cloud framework necessary to allow for TSoS monitoring and control; Formulation of a data
exchange model to ensure successful information transmission and understanding
• Task 2.3
• Goal: Identification of requirements for both the control theoretical and software tools from
the perspective of practical applications via two real-world TSoS cases
• Outcomes: Identification and definition of specific objectives, capabilities, constraints and
requirements concerning the implementation of the Local4Global control approach to:
• A building TSoS use case concerning the E.ON ERC main building in Aachen, Germany
• A traffic TSoS use case concerning a signal-controlled road section of more than 5 km in Munich,
Germany
9
WP2: L4G-to-SoS Requirements
Timeframe and work allocation
• Timeframe
• Start:
October 1, 2013
• Duration: 9 months
• End:
June 30, 2014
• Work allocation to partners and deliverables
WP2 Leader: TUC
2.1
TUC
CERTH, ETHZ
D2.1
2.2
IK4
CERTH, ETHZ
D2.2
2.3
TRV (traffic use case)
RWTH (building use case)
CERTH, ETHZ, TUC, TUM
D2.3
10
WP3: TSoS Analysis and Modelling Tools
Work structure and outcomes
• Task 3.1
• Goal: "Just Enough" Learning for TSoS: develop and successfully test (through simulations) the selflearning mechanism of Local4Global.
• Outcomes: System identification and learning approaches for Hybrid Systems combined together
with machine learning approaches for extracting the necessary information from large amount of
information data towards such a purpose. The self-learning mechanism will be extensively tested
using accurate and quite realistic simulator models for the two Local4Global Use Cases.
• Task 3.2
• Goal: "Just Enough" Situation Awareness for TSoS: develop and successfully test (through
simulations) the situation awareness mechanism of Local4Global.
• Outcomes: State observers and filters for Hybrid Systems will be combined together with machine
learning approaches for extracting the necessary information from large amount of information
data towards such a purpose. The self-learning mechanism will be extensively tested using accurate
and quite realistic simulator models for the two Local4Global Use Cases.
• Task 3.3
• Goal: Optimality, Certification and Stability in TSoS: extend notions from standard control theory
such as stability, clost-to-optimality, etc to TSoS and develop a mechanism for determining when
action by the optimization algorithm is needed in the TSoS context
• Outcomes:
• Extension of notions of optimal control, Lyapunov functions, certification, etc to hybrid systems and
networked control systems developed by Local4Global partners in the past will serve as the basis towards such
a purpose
• Development of a methodology of event-based optimization, where the optimization algorithms interacting
with the low level controllers take the “back seat” and intervene only whenever necessary. Motivated by the
Use Case studies, investigate criteria for generating the events to trigger the actions of the optimization
algorithms. The aim will be to minimize the communication and coordination necessary but still preserve
fundamental system theoretic properties of the closed loop system, in particular ensure stability and provide
bounds on the sub-optimality of the resulting solution when compared with running the optimization
continuously (as is done, for example, in traditional model predictive control).
11
WP3: TSoS Analysis and Modelling Tools
Timeframe and work allocation
• Timeframe
• Start:
October 1, 2013
• Duration: 12 months
• End:
September 30, 2014
• Work allocation to partners and deliverables
WP3 Leader: ETHZ
3.1
CERTH
TUC
D3.1
3.2
CERTH
TUC
D3.2
3.3
ETHZ
CERTH, TUC
D3.3
12
WP4: The Local4Global System
Work structure and outcomes
• Task 4.1
• Goal: TSoS Distributed Optimization: Development of a distributed optimization module to accompany and
cooperate with the Self-Learning and Situation Awareness Mechanisms.
• Outcomes: Real-time decision making methods for generating decisions in a worst-case or robust manner,
based on randomized real-time decision making.
• Task 4.2
• Goal: C4L/L2C and the Local4Global System: Development of the C4L/L2C components and of the overall
Local4Global system (methodology)
• Outcomes: Deliver the C4L/L2C mechanism of Local4Global as well as the overall methodology for the
Local4Global system fulfilling the following objectives:
• Fully scalable and computationally efficient implementation.
• Efficiently perform locally and without the assumption of perfect and complete knowledge about the
overall TSoS state and dynamics.
• Identifying and predicting the emerging and evolutionary characteristics at the macro-level and
controlling the overall TSoS dynamics (moving in the "right direction“).
• Controlling the emergent and evolutionary behaviour of TSoS by adaptively assigning different
objectives and tasks to different TSoS subsystems.
• Task 4.3
• Goal: The Local4Global Integrated Software System: Integration of all different Local4Global components and
mechanisms and delivery of the final ready-to-be-deployed product.
• Outcomes: Local4Globlal middleware and the Local4Global cloud services: Middleware will be the glue
between huge amounts of connected things and Cloud services will offer a complete set of functionalities to
the world.
13
WP4: The Local4Global System
Timeframe and work allocation
• Timeframe
• Start:
July 1, 2014
• Duration: 14 months
• End:
September 30, 2015 (currently active)
• Work allocation to partners and deliverables
WP4 Leader: CERTH
4.1
ETHZ
CERTH, TUC
D4.1.1/D 4.1.2
4.2
CERTH
TUC
D4.2.1/D4.2.2
4.3
IK4
CERTH, TUC
D4.3
14
Task 4.1: TSoS Distributed Optimization
• Based on outcomes from WP3
• Distributed optimization mechanisms were designed and implemented
using only available local data and a global index indicating the efficiency
at the global level.
• Together with the Self-learning mechanisms and Situation Awareness
mechanisms, a local control generation mechanism was integrated forming
the distributed optimization process of the L4G system.
15
Task 4.2: C4L/L2C and the Local4Global System
• Local Optimization for Global
Performance
Mechanism:
Incorporates the main core of the
adaptive/learning attributes of
the
Local4Global
control
approach, integrating the whole
idea of distributed optimal control
towards achieving global level
control efficiency.
Through random control strategy
noise inserted by the random
control perturbations persistent
excitation of each constituent
system is achieved.
Through efficient self-learning and
situation-awareness
modules,
convergence towards maximizing
global control performance is
achieved.
More details will be given in the L4G
system presentation later on.
16
Task 4.3: The L4G Integrated Software System
• Generic interface design:
• Common functionalities and
needs for all the controllers
• Different implementations for
different controllers
• Architecture methodology flow:
• Libraries are used for algorithm
validation in simulation
environments
• Reuse libraries for the integrated
software system
• First implementation:
• Traffic use case controller
17
Task 4.3: Controller Implementation (traffic)
• Library integration:
• TUC.dll and L4GCAO.exe
• Successful runs with mock data
• Adaptability:
• Configuration files can be used to adapt the controller to different traffic
contexts (junctions, plans, detectors...)
• Code modularity reduces complexity of future updates
• Work to be done:
• Real data inputs & outputs (real-life implementation):
• Develop additional access protocols
• Additional simulation tests for evaluation:
• Use same data to compare results with simulation
• Improve control logic design using information form driving cars
• Integrate updated algorithm libraries
• Design & develop building controller implementation
18
WP5: Use Cases
Work structure and outcomes
• Task 5.3 (Transver/TUM)
• Goal: Implementation, verification and operation of the Traffic Use Case
• Outcomes: Final implementation plan D5.1.2 defining the use case
architecture, the implementation steps and the risk assessment, considering
the test bed technical, organizational and institutional environment; creation
of microscopic traffic simulation to test the control strategy and pre-adjust
environment parameters to help in the test bed installation, and to perform
evaluations based on artificial scenarios
• Task 5.4 (RWTH)
• Goal: Implementation, verification and operation of the Building Use Case
• Outcomes: Final implementation plan D5.2.2 defining the integration
strategy within the test bed‘s control environment as well as the simulation
model of the test bed for simulative pre-analysis and artificial test scenarios;
Run capable co-simulation applying the first version of the Local4Global
algorithm in conjunction with a physical model of the building TSoS;
simulative baseline calculation; base case simulation ready; test case
simulation in final adjustment
19
WP5: Use Cases
Timeframe and work allocation
• Timeframe
• Start:
March 1, 2014
• Duration: 27 months
• End:
June 30, 2016 (currently active)
• Work allocation to partners and deliverables
WP5 Leader: Transver
5.1
Transver
TUM, TUC
D5.3
5.2
RWTH
CERTH, IK4
D5.4
5.3
Transver
TUM, TUC
D5.1.1/D5.1.2
D5.5
5.4
RWTH
CERTH
D5.2.1/D5.2.2
D5.6
20
WP6: Evaluation
Work structure and outcomes
• Task 6.1 (RWTH/Transver/TUM)
• Goal: Produce the evaluation plan
• Outcomes: First and final evaluation plans D6.1.1 and D6.1.2 were
produced, to prescribe the general evaluation approach, the assessment
objectives, the indicators to be used, the methods and timing of
measurement and the measurement conditions, and statistical issues such as
sample sizes. The plan describes the project’s KPI, and defines the
comparison baseline of each use case.
• Task 6.2 (RWTH/Transver/TUM)
• Goal: Evaluation of each use case
• Outcomes: Although this task is intended to start at M27, some preparation
work has already been anticipated. Namely, computer simulations have
been prepared to make evaluations of artificial scenarios for both use cases.
21
WP6: Evaluation
Timeframe and work allocation
• Timeframe
• Start:
October 1, 2013
• Duration: 36 months
• End:
30 September, 2016 (currently active)
• Work allocation to partners and deliverables
WP5 Leader: Transver
6.1
RWTH
TUM, Transver
D6.1.1/D6.1.2
6.2
RWTH
ALL PARTNERS
D6.2
6.3
IK4
ALL PARTNERS
D6.3
22
WP7: Dissemination, Standardization, Exploitation
and Business Plan
Work structure and outcomes
• Task 7.1 (CERTH; RWTH/Transver/TUM)
• Goal: Dissemination
• Outcomes: Website, brochures (for L4G and L4G products), videos, press
release, journal papers, conferences
• Task 7.2: Business Model, Standardization
(IK4/CERTH/ETHZ/RWTH/TRV/TUC/TUM )
&
Exploitation
• Goal: Define partners exploitation and commercialization strategy of L4G
project results.
• Outcomes: First version of exploitation deliverable was produced,
describing exploitable products and each partner’s strategy to market
them. In addition first version of L4G Platform, L4G Energy Efficiency Tool and
L4G Mobility Tool business models has been elaborated. A set of standards
have been selected covering different areas in Local4Global system, also
describing their implementation in the L4G Integrated Software System.
23
WP7: Dissemination, Standardization, Exploitation
and Business Plan
Timeframe and work allocation
• Timeframe
• Start:
October 1, 2013
• Duration: 36 months
• End:
30 September, 2016 (currently active)
• Work allocation to partners and deliverables
WP7 Leader: IK4
7.1
CERTH
IK4, ETHZ, RWTH, TRV,
TUC, TUM
D7.1/D7.2.1
D7.3.1/D7.3.2
7.2
IK4
CERTH, ETHZ, RWTH,
TRV, TUC, TUM
D7.4
D7.5.1/D.7.6
24
Thank you!
25