IEEE Technology and Society Magazine - Fall 2014 - 39

would use detectors at each of the
points (A, B, C, and D), and define
two patterns consisting of the corresponding sequences. The system
would detect these sequences using
finite state automata. The volume
and velocity of the events that are
required in order to be processed,
as well as the complexity of some
of the automata, require distributing the automata-processing task
among multiple nodes.
A more realistic solution to the
path-counting task is to take into
account the uncertainty in the detection of the locations of vehicles.
Detectors may fail to detect some
vehicles, may have false detections,
and may report detections that are
inherently uncertain (e.g., locating vehicles via a cellular network).
Automata used for detecting patterns over deterministic events are
unsuitable in this scenario. On the
other hand, as discussed previously,
probabilistic models such as Markov Logic Networks are designed to
handle uncertainty, and are therefore
a natural choice for detecting events
under uncertainty. Event recognition
and forecasting with Markov Logic
Networks can be done by inference
over probabilistic graphical models,
which is fundamentally different
from computations over state automata. Consequently, distributing these
tasks among multiple nodes requires
correspondingly different algorithms.
To address the Big Data issues
of volume, velocity, and lack of
veracity, methods are required for
distributing event recognition and
forecasting tasks that incorporate
probabilistic reasoning. This requires
distributing on-line inference tasks
among multiple nodes, as opposed to
state automata used for recognition
tasks over deterministic events. Such
algorithms can exploit the continuous nature of the recognition task by
incrementally modifying the inference as new events occur.
In addition to the computational
scalability issues, the increasing number of distributed event-generating
sources requires that inherently limited

network resources be employed efficiently. For example, in traffic management some sensors may be deployed
at locations where a high-speed wide
area network is not available, and will
therefore be required to continuously
transmit a high volume of sensor
readings via a cellular network. Since
communication efficiency reduces the
volume of data sent to a data center for
processing, it may also improve computational efficiency. Communication
efficiency also helps in maintaining
the privacy of the entities generating
the events (e.g., terminals in credit
card transactions).
Communication-efficient distributed detection has been an active
research field in recent years. Proposed methods reduce communication by decomposing the recognition
task into a set of local constraints
on the data generated at the nodes.
The constraints are such that, as
long as all of them are upheld, it is
guaranteed that the event of interest
has not occurred. Consequently, as
long as all constraints are upheld,
no communication is required. The
event to be recognized is usually
defined using a function over aggregate values derived at the nodes. In
other words, event recognition is
restricted to numerical reasoning.
To support the full range of
functionality required by Big Data
applications, the development of
distributed communication-efficient
event recognition, and forecasting algorithms is necessary. This
includes events defined over aggregates, as well as temporal, logical,
and spatial patterns over events.
Emphasis needs to be placed on
handling functions that do not have
a closed form, such as inference
over probabilistic graphical models.

Event-Driven
Decision Making
In the proposed methodology, the
forecast events, along with the recognized events, are leveraged for
real-time operational decision making. A body of tools for real-time
proactive decision making exploits

IEEE TECHNOLOGY AND SOCIETY MAGAZINE

FALL 2014

the event forecasting models presented above, with an emphasis on
optimization methods that intelligently handle forecast uncertainty
using robust, stochastic, or blackbox methods.
In terms of real-time optimization techniques, the state-of-the-art
is that optimization techniques are
being activated mostly off-line and
use a variety of methods that fit
different assumptions, e.g., robust
(worst-case) optimization or stochastic optimization. In the field of robust
optimization methods, the state-ofthe-art focuses on tools for providing strong performance guarantees
for convex optimization problems
[3]. For real-time decision-making
purposes, the use of robust optimization methods involving recourse,
that is, modeling the notion that
future decisions can be deferred until
future information is available, is an
area of intensive ongoing research.
For example, in the context of traffic management, "recourse" decisions refer specifically to traffic
management actions. This includes
the alteration of speed limits and
restriction of on-ramp flows, which
are computed as future responses to
changes in traffic flows, and result
from similar actions taken at an early
time. The use of "robust" or "worstcase" models is most appropriate for
those aspects of traffic management
with hard limits, such as absolute
limits on allowed flows or maximum
closure time constraints.
Stochastic, or randomly determined, optimization focuses on optimizing an expected value criterion that
is subject to probabilistic constraints.
Aside from the need to parameterize policies in the recourse sense discussed above, an additional difficulty
relates to the interpretation of constraints. Due to the probabilistic nature
of the uncertainty that enters the optimization, the worst-case constraints
used in robust optimization often turn
out to be impractical. One then has to
resort to soft interpretations, such as
chance constraints, ensuring that the
probability of meeting the constraint
|

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - Fall 2014