A Complex Event

boorishadamantΤεχνίτη Νοημοσύνη και Ρομποτική

29 Οκτ 2013 (πριν από 3 χρόνια και 7 μήνες)

104 εμφανίσεις

A Complex Event
Recognition Architecture

Will Fitzgerald

Kalamazoo College


R. James Firby

I/NET, Inc
.

A Complex Event
Recognition
Architecture

Protecting us from
the Metal Horde!

Will Fitzgerald

R. James Firby

What is …


A Complex Event?


Complex events are hierarchical,
discrete, time
-
stamped structures
inferred from multi
-
channel,
asynchronous signals.

What is …


A Complex Event
Recognition Architecture?


A description or
implementation of typical
patterns and recognition
algorithms for complex
events.


A real example…


Water Recovery
System at NASA's
Johnson Space
Center


Four complex
subsystems,



About 200 sensors
and actuators,


Each subsystem
asynchronously
signals data.

Detecting “Safe Mode”


When a problem is detected
internally, the Water Recovery
System attempts to go into
“safe mode,” which occurs
when the four subsystems are
safed.


“Safing” of the four subsystems
happen asynchronously.


“Safing” detection for each
subsystem differs from one
another.


On recognizing that the WRS
has gone into safe mode, signal
an event that all subsystems
have been safed.

Another example…

To get directions to a
location on the on
-
board map, the user
says:


“Go here” and



Taps the display
location


within 200 ms.


(CNN photo)

Parsing the world


Dynamic Predictive Memory Architecture
(DPMA)


KR and Semantic parsing


Task execution and dialogue management



“complex, dynamic environments”


Do similar techniques apply to …


multi
-
channel, asynchronous sensors?


multi
-
modal interface input?

A Complex Event Recognition
Architecture


What
assumptions

are reasonable to
make about the form of input data?


What
useful general patterns

are
there in the data?


What
recognition algorithms
do we
need?

NLP Assumptions


Input to Natural Language Processing
systems are typically assumed to be:


Discrete events of one type (“words”)


Single channel


Totally ordered by position; duration
irrelevant


1

2

3

4

5

time

flies

like

an

arrow

More generally…


Events of various types


Over multiple channels and asynchronous


Duration of event often important


“Hierarchical” model still useful

000

005

010

015

020

025

030

put

these

here



click and drag


tap

Assumptions about Events


Discrete:
Individually distinict, non
-
continuous data (could be discretized).


Time
-
stamped:

Event carries the start
and end times (defining the event
duration
, which could be instanteneous).


Typed:

Events form distinct types (e.g.,
words

vs.
taps
).


Structured:

Event may internal,
hierarchical structure (“complex”).

Standard Event Patterns


Are there patterns of events which are
particularly useful to identify?


Are there recognition algorithms to identify those
patterns?


Yes.


ONE and BINDING


IN
-
ORDER, ALL, ONE
-
OF


Allen patterns


WITHIN and WITHOUT


ONE

and
BINDING

patterns


ONE
: The simple
pattern of looking for
a single event (of a
particular type).


BINDING:

ONE

pattern plus collecting
and constraining
state.


Essentially “event
-
driven” programming;
the “stimulus” in S
-
R.


ON
-
CLICK


A
ONE

pattern if just
looking for the ‘click’


A
BINDING

pattern if
x,y

coordinates are
significant.

IN
-
ORDER

patterns


Events will occur “in order”


That is, saying two events,
A

and
B
, occur in
order, the
start time

of B is


the
end time

of
A.


(IN
-
ORDER A B C D)


First an event of type A, then B, etc.

IN
-
ORDER

as
NLP


Combined with
BINDING

and signaling of
subpatterns this is essentially a classic
natural language processing pattern.

S


NP VP

NP


DET N

VP


V NP

The boy saw the girl.

[S [NP [DET the][N boy]]


[VP [V saw]


[NP [DET the] [N girl]]]]

ALL
Patterns


Events will all occur, but in any order


With this, we leave (our) standard NLP
approaches.


For example, user will choose from all of the
sets of options.


For example, all subsystems will be “safed”,
but in any order.

ALL patterns and
contradiction


The problem: user or system “undoing” an
event that has already been seen
(interpreting events as state changes).


Example: Class will start when all the
students, Alice, Bob, Charles, Dominique,
have arrived.

Consider this sequence for

(ALL A B C D):

1.
Charles arrives.

2.
Alice and Bob arrive together.

3.
Alice starts to sing.

4.
Charles leaves.

5.
Dominique arrives.

6.
Charles arrives.

Order is not relevant; Alice’s singing is not relevant;
but Charles’s leaving undoes his earlier arrival.

ONE
-
OF Pattern


Look for any of a set of event forms


Example: Office hours begin as soon as
one of the professors A,B,C or D arrives.


(ONE
-
OF A B C D)

Time
-
based patterns


Allen relationships


WITHIN

patterns


WITHOUT

patterns


Allen Patterns


James Allen
described the
relationships between
two intervals.


Allen patterns look for
temporal relationships
between 2 events or
an event and an
interval.

1.
contains

2.
finishes

3.
starts

4.
before

5.
meets

6.
overlaps

7.
equal

8.
overlapped by

9.
after

10.
met by

11.
started by

12.
finished by

13.
during

A

contains

B

A




overlaps
B

WITHIN and WITHOUT


WITHIN
patterns
reflect that the
duration of an event is
no longer than a
certain amount of
time.


E.g.
, an
ALL

pattern
wrapped in a
WITHIN

pattern.


WITHOUT

patterns
reflect that an interval
of time will pass
without the
occurrence of an
event.


E.g.
, Sherlock
Holmes’s
“significance of the
barking dog.”

Pattern Combination


“Go here” and a tap
within 200 ms.


(within


(all


(in
-
order go here)


(tap ?x ?y))


200 ms)

(CNN photo)

Safe mode recognizer

(define
-
recognizer (safing
-
complete)


(pattern


'(all


(safing (system pbbwp) (status on))


(safing (system ro) (status on))


(safing (system aes) (status on))


(safing (system pps) (status on))))


(on
-
complete (st end)


(signal
-
event '(all
-
safed) st end)))

Some details elided…

Parsing Algorithms


The parsing algorithms and recognizer
semantics are more fully described in the
paper.


Implementation Details

Conclusions


Standard patterns of events.


Standard recognizers for these patterns.


Good for monitoring complex (internal)
system state.


Useful for recognizing patterns of complex
events over multiple modes, over time.

Acknowledgments


Work done under NASA SBIR contract
NAS9
-
00122.


We would like to especially acknowledge
collaborators at NASA, including Debra
Schreckenghost, Pete Bonasso, Carrol
Thronesbery and others.


Pulp Images from “Pulp of the Day”:

groups.yahoo.com/group/pulpoftheday


Questions?