-
International ConferenceNANOSENSORY SYSTEMS AND QUANTUM SENSORY
MATERIALSD. Metafas1, M. Rangoussi1, B. Meparishvili2, G.
Goderdzishvili2
(1) Department of Electronics Eng., TEI Piraeus, Greece(2)
Faculty of Informatics and Management Systems, Georgian Technical
University, Georgia
Dynamic Applications on Complex Distributed Systems and Machine
Learning algorithmsInternational ConferenceNANOSENSORY SYSTEMS AND
QUANTUM SENSORY MATERIALS Tbilisi, Georgia 6-9 Jun 2013
-
Internet of Things (IoT)
IoT: Focus on Critical Infrastructures
SCADA and Wireless Sensors and Actuator Networks (WSANs)
Review of existing middleware for networked robots and Wireless
Sensor Networks
The need of a WSAN middleware
WSAN middleware: Focus on autonomous decision-making
Machine LearningOutlineInternational ConferenceNANOSENSORY
SYSTEMS AND QUANTUM SENSORY MATERIALS
-
Internet of Things Internet of Things (IoT) the next step in
always on computing promising a world of networked and
interconnected devices. ITU 2005Machine-to-Machine, smart systems,
web of things, sensor web, The real (i.e., physical), virtual, and
digital worlds are converging, thanks to an ever proliferation of
connected (smart) sensors and objects, ubiquitous wireless
networks, communications standards and the activities of humans
themselves Economist 2010
International ConferenceNANOSENSORY SYSTEMS AND QUANTUM SENSORY
MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Internet of ThingsThe vision of IoT is changing from emerging
realizations, via e.g., RFIDs and WSANs, to include
Internet-connected objects, whereby objects will include any
possible physical entity that could be classified according to
size, mobility, power, connectivity, automation, physical/logical
type, etc. Internet-connected objects is a central part within
International/European strategies/roadmaps. This change is
characterized as metamorphosis of objects in EU Cluster of European
Research Projects on the IoT. A key enabler/prerequisite for the
proliferation of internet-connected objects is the widespread
deployment of IPv6.
International ConferenceNANOSENSORY SYSTEMS AND QUANTUM SENSORY
MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
IoT: Focus on Critical InfrastructuresInternational
ConferenceNANOSENSORY SYSTEMS AND QUANTUM SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
IoT: Focus on Critical InfrastructuresOur main focus is Critical
Infrastructures (CIs), i.e., those facilities that provide
essential services for everyday life (such as energy, food, water,
transport, communications, health).
Protecting Critical Infrastructures (CIs) against disruption of
any kind is deemed vital to the national/international security,
public health and safety, as well as economic prosperity.PCS
(Process Control Systems)/SCADA (Supervisory Control and Data
Acquisition) systems are largely used for data acquisition and
control over large and geographically distributed infrastructures;
these systems are critical elements in every aspect of literally
every CI (e.g., Oil/Gas, Electricity, Transportation).
International ConferenceNANOSENSORY SYSTEMS AND QUANTUM SENSORY
MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Critical Infrastructures: SCADA systemsSCADA systems range from
relatively simple networks that monitor environmental conditions of
a given location to incredibly complex systems that monitor all the
activity in a power plant or a municipal water system.
International ConferenceNANOSENSORY SYSTEMS AND QUANTUM SENSORY
MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
SCADA and Wireless Sensors and Actuator Networks (WSANs)SCADA
systems are moving from expensive, complex, proprietary
technologies to affordable, standards-based, open solutions. One of
the promising (in cost and usage) technologies for use in SCADA
systems includes IEEE 802.15.4-based Wireless Sensors and Actuator
Networks (WSANs) such as ZigBee, WirelessHART and ISA100.11a.
but ..
Wireless systems are vulnerable to in terms of cyber attacks.CIs
(like oil and gas and power grid) are attractive target for
cyber-attacks.The use of WSANs in CIs results in a deviation from
the WSAN assumption of densely deployed nodes. This is not the case
when they are used for SCADA of CIs, and a non-communicating node
can have much more significant effect than the standard WSAN
scenario.International ConferenceNANOSENSORY SYSTEMS AND QUANTUM
SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
WSAN middlewareMost current middleware solutions for WSAN
systems fail to take into account two key aspects: robustness of
the solution and support for heterogeneous configurations. This is
especially important in environments where their complexity makes
it absolutely necessary to have different types of techniques and
monitoring equipment that needs to interact to project CIs.
For example, the case where oil pipelines are to be monitored
require the cooperation and interaction of sensors, actuators and
mobile devices (UAVs, robots, etc.) that interact in a seamless way
to achieve a common goal. International ConferenceNANOSENSORY
SYSTEMS AND QUANTUM SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Dynamic Applications for CIs and the need of a high level
middlewareDynamic applications for CIs require complex distributed
systems consisting of a number of embedded hardware and software
components. Since these components are heterogeneous, it is
necessary to find ways to mask this heterogeneity and offer some
innovative solutions to develop the applications. In addition,
advanced and efficient techniques for cooperation and collaboration
among them are required to achieve the desired goals. Therefore,
there must be new software services (middleware) that act as the
glue to link everything together in an efficient manner, supporting
concurrency-intensive operations, enhancing collaboration, and
insuring efficiency and robustness. How to get efficient
communication links, design better ways to exchange and use
information over these links and manage dynamic work environments
are some of the common issues.Proposed middleware approaches for
specific technologies such as ad-hoc networks, wireless sensor
networks and robotics are addressing a number of issues but clearly
in the broader scope of Dynamic applications for CIs there is a
need of a widely accepted high-level middleware addressing all open
issues and meet the design and implementation of different
challenges. International ConferenceNANOSENSORY SYSTEMS AND QUANTUM
SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Dynamic Applications for CIs and the need of a high level
middleware (2)Objective: To design a distributed middleware and
communications framework that will equip objects/nodes with
advanced functionalities and abilities to understand their
environment for collaborative and autonomous decision-making. To
prescribe all the components of the distributed middleware, which
shall enable cooperation, intelligence and autonomy among
heterogeneous, spatially distributed nodes/objects, while providing
self-X capabilities (self-configuration, self-healing,
self-management, self-optimization).To develop a mission-to-system
translation methodology allowing for a systematic decomposition of
high-level mission requirements, to low-level system primitives and
parameters. The methodology will allow for reusability across
various application areas, potentially reducing the effort and the
development time.International ConferenceNANOSENSORY SYSTEMS AND
QUANTUM SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Reviewing the middleware for Networked Robotsand WSNsNetworked
Robots middleware:Concerning the communication model: MIRO, PEIS
Kernel, Orca (CORBA but also non standard models)Concerning the
Interoperability of the software components: Marie, Middle
LayerConcerning the Automatic Discovery, Configuration and
Integration support: PEIS Kernel, UPnPConcerning the
Specific/Expandable Services: Sensory Data Processing Middleware,
the AWARE data-centric model, MIRO, MarieConcerning the
communication performance issues (reliability, availability or
QoS): RSCA (Robot Software Communication Architecture)Concerning
the support of Low Resources devices: PEIS KernelWSN middleware:
Moteview, ScatterViewer, Hourglass, SenseWeb, jWebDust, GSN,
TinyDB, Hood, SNACK, Kairos, International ConferenceNANOSENSORY
SYSTEMS AND QUANTUM SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
WSAN middleware: Focus on autonomous decision-makingGiven the
model of the WSAN system ..Train the system using different
scenarios, conclude to AI brains and provide them to the AI agents.
An innovative idea we are exploring is the use of Q-Learning
algorithms and model the Q as a Model Tree.
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Machine LearningClassification:GivenTraining dataLearnA model
for making a single prediction or decisionInternational
ConferenceNANOSENSORY SYSTEMS AND QUANTUM SENSORY MATERIALSSource:
Lisa Torrey, Univ. of Wisconsin Madison (2009)
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Machine LearningProcedural Learning:Learning how to act to
accomplish goalsGivenEnvironment that contains rewardsLearnA policy
for acting
Important differences from classificationYou dont get examples
of correct answersYou have to try things in order to
learnInternational ConferenceNANOSENSORY SYSTEMS AND QUANTUM
SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Machine LearningWe dont know all the effects of our decisions
so, the procedural learning drive us to the Reinforcement Learning
(RL) (acting and observing the environment - or the model of it in
our case - ).What kind of RL ? Model-free: learn a policy without a
strict model. Use Temporal difference methods (TD).What kind of
Model-free RL ? Q-Learning
International ConferenceNANOSENSORY SYSTEMS AND QUANTUM SENSORY
MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Q-LearningCurrent state: sCurrent action: a
Transition function: (s, a) = s
Reward function: r(s, a) R
Policy (s) = a
Q(s, a) value of taking action a from state sInternational
ConferenceNANOSENSORY SYSTEMS AND QUANTUM SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Q(s, a)The Q (state, action) table is huge ..
What we could do ? Approximate the Q with a non-linear
function.
What kind of function ? A Neural Net or a Model Tree.
Lets test the Model Tree. If its application is successful we
could have a much more comprehensive and flexible set of rules for
our AI agents.
Model trees are similar to decision trees, but instead of
predicting discrete classes, they predict real valued functions.
Model trees recursively partition the feature-space into regions,
by choosing one attribute as a split criterion at every level of
the tree. In the leaves of the trees, regression models are trained
to predict a numerical value from the features of an instance.
What algorithm ?
A variant of the Quinlan's M5 algorithm (Quinlan 1992). This
algorithm first grows a tree by selecting splitting criteria so as
to minimize the target variable's variance, and then builds linear
regression models in the nodes. Finally the tree is pruned, which
means that sub-trees are replaced with leaves, as long as the
prediction error of the resulting tree does not exceed a certain
threshold.
International ConferenceNANOSENSORY SYSTEMS AND QUANTUM SENSORY
MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Decision Trees (ID3)International ConferenceNANOSENSORY SYSTEMS
AND QUANTUM SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
M5 algorithmExampleInternational ConferenceNANOSENSORY SYSTEMS
AND QUANTUM SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
The problems ..The main disadvantage in using model trees for
value function approximation is that there is currently no
algorithm for online training. To refine the predictions of a model
tree, we must therefore rebuild it from scratch, using not only the
new training examples, but also the old ones, that were used for
the previous model trees, i.e. a new model tree approximation of
the Q-functions (more than one because of the hierarchical model
trees) has to be built from the whole updated training set, and
these trees are then used to play the next set of training data.
Even though there is not experience with using model trees in
reinforcement learning we believe that it is a promising
approach.International ConferenceNANOSENSORY SYSTEMS AND QUANTUM
SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Next steps ..To prescribe the components of the distributed
middleware, which shall enable cooperation, intelligence and
autonomy among heterogeneous, spatially distributed
nodes/objects.To develop a mission-to-system translation
methodology allowing for a systematic decomposition of high-level
mission requirements, to low-level system primitives and
parameters.If the experiments of the proposed approach are
successful, the methodology could be based on hierarchical Model
Trees generated AI agents, trained for a number of
scenarios.International ConferenceNANOSENSORY SYSTEMS AND QUANTUM
SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
-
Thank you!International ConferenceNANOSENSORY SYSTEMS AND
QUANTUM SENSORY MATERIALS
Master of Science in Embedded Systems DesignALaRI Faculty of
informatics - University of Lugano, USI
****