D4.1 – First Synopsis of the RUBICON Self Organising Fuzzy Neural ...

prudencewooshAI and Robotics

Oct 19, 2013 (3 years and 11 months ago)

136 views

 
RUBICON
 
Robotic  UB
I
quitous  COgnitive  Network
 
 
 
 
 
Project  No.:  
269914
 
 
 
Dissemination  level
 
X
 
PU
 
=  Public
 
 
PP
 
=  Restricted  to  other  programme  participants  (including  the  Commission  Services)
 
 
RE
 
=  Restricted  to  a  group  specified  by  the  consortium  (including  the  Commission  Services)
 
 
CO
 
=  Confidential,  only  for  members  of  the  consortium  (including  the  Commission  Services)
 
 
©  Copyright  
RUBICON
 
-­‐
 
All  Rights  Reserved
 
D
4
.1


First Synopsis of the
RUBICON Self Organising
Fuzzy Neural Network
(SOFNN)

 
 
Editor:
 
T.  M.  
McGinnity
 
ULSTER
 
 
Contributor(s):
 
Anjan  Kumar  Ray
 
ULSTER
 
 
Liam  Maguire
 
ULSTER
 
 
Sonya  
Coleman
 
ULSTER
 
 
Mauro  Dragone
 
UCD
 
 
Mathias  Broxvall
 
ORU
 
 
Claudio  Gallicchio
 
UNIPI
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
2
 
 
 
Issue  Date
 
30
/
09
/
2011
 
(M9,  MS1)
 
Deliverable  Number
 
D
4
.
1
 
WP
 
WP4
 
-­‐
 
Cognitive
 
Layer
 
Status
 
¨
Draft  
¨
Working  
x
Released  
¨
Delivered
 
to  EC
 
¨
Approved  by  EC
 
 
 
 
 
Document  history
 
V
 
Date
 
Author
 
Description
 
1
 
30
/9
/2011
 
UU
 
D4.1  Final
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
3
 
 
Disclaimer
 
The  information  in  this  document  is  provided  as  is  and  no  guarantee  or  warranty  is  given  
that  the  information  is  fit  for  any  particular  purpose.    The  user  thereof  uses  the  information  
at  its  sole  risk  and  liability.  
 
The  document  reflects  only  the  author’s
 
views  and  the  Community  is  not  liable  for  any  use  
that  may  be  made  of  the  information  contained  therein.
 
 
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
4
 
 
Executive  
Summary
 
Th
is
 
deliverable
,
 
D4.1
,
 
outlines  the  ‘first  synopsis  of  the  RUBICON  Self  Organising  Fuzzy  
Neural  Network  (SOFNN)’  as  part  of  the  development  of  the  cognitive  layer  (WP4)  for  the  
RUBICON
 
project.  
The  overall  goal  of  this  work
-­‐
package  is  to  develop  cognitive  mechanisms  
to  support  
the  RUBICON  ecology.  
 
In  Chapter  
1,  we  present  the  overall  high
-­‐
l
evel  
architecture  
of  the  
RUBICON
 
ecology
,  as  
agreed  within  the  consortium
.  We  highlight  the  connections  of  the  cognitive  layer  with  the  
learning,  control  and  communication  layers.  We  briefly  
describe  the  overall  role  of  the  
cognitive  layer
 
in  meeting  the  objectives  of  the  project.
 
Chapter  
2  reviews  the  requirements  of  a  general  cognitive  architecture.  We  present  selected  
cognitive  architectures
,
 
w
hich  
have  been  reported  
in  the  literature
 
over  
recent  decades
 
and  
consider  the  most  appropriate  for  RUBICON.
 
 
In  Chapter  3,  we  present  the  structure  of  the  Self  Organising  Fuzzy  Neural  Network  (SOFNN),  
which  we  anticipate  will  form  the  basis  for  the  design  of  the  cognitive  layer.  Applications  of  
the  SOFNN  
for  function  approximation,  system  identification,  and  time
-­‐
series  prediction  are  
presented.  The  challenges  and  possible  usage  of  the  SOFNN  structure  within  the  RUBICON  
project  are  also  considered.
 
 
In  
Chapter  4
,
 
we  define
 
the  requirements  of  the  cognitive  architecture  within  the  
RUBICON
 
project.  These  requirement
s  
are  
designed  to  embody
 
requirements  of  D2.1,  D3.1  and  
inputs  
from  several  discussions  within  the  project  consortium.
 
In  Chapter  5
,  we  discuss  
in  more  detail  
the  interfaces  
between
 
the  cognitive  layer  
and  
the  
learning  and  control  layers.  
Firstly,  w
e  prese
nt  the  data  structure  for  communication  
between  the  learning  
and  cognitive  layers  and  specify  a  strategy
 
to  decode  distributed  
location  information  from  the  data  arriv
ing  
from  the  learning  layer.  
Secondly,  w
e  present  the  
data  structure  for  communication  between  the  cognitive  and  control  layers  and  present  a  
parameterised  approach  to  handle  
an  
increasing  number  of  control  goals  within  the  
RUBICON
 
ecology.
     
 
Following  the  discussions  of  the  interfaces  between  
major  component  blocks  in  
RUBICON
,  
w
e  present  
further  detail  of  the
 
high  level  design  of  the  
actual  
cognitive  architecture  for  the  
RUBICON
 
ecology  in  Chapter  
6
.  We  categor
ise
 
the  architecture  into  three  main  modules:  
cognitive  memory,  cognitive  reasoning  a
nd  cognitive  decisions.  We  discuss  the  
interconnections  and  functioning  of  these  modules  from  the  perspective  of  the  
RUBICON
 
project.      
 
Chapter  7  concludes  the  deliverable  with  a  short  summary  of  the  forward  plan  and  major  
aspects  to  be  investigated.
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
5
 
 
Cont
ents
 
EXECUTIVE  SUMMARY
 
................................
................................
................................
............
 
4
 
ABBREVIATIONS
 
................................
................................
................................
......................
 
7
 
FIGURES
 
................................
................................
................................
................................
..
 
9
 
TABLES
 
................................
................................
................................
................................
..
 
10
 
1.
 
INTRODUCTION
 
................................
................................
................................
.............
 
11
 
1.1
 
O
VERVIEW  OF  THE  
H
IGH  
L
EVEL  
RUBICON
 
A
RCHITECTURE
 
................................
................................
 
11
 
1.2
 
R
OLE  OF  THE  
C
OGNITIVE  
L
AYER  IN  THE  
RUBICON
 
PROJECT
 
................................
...............................
 
12
 
2
 
OVERVIEW  OF  EXISTING
 
COGNITIVE  SYSTEMS  AR
CHITECTURES  IN  THE  L
ITERATURE
 
..
 
15
 
2.1
 
A
PPROACHES  TO  
A
RTIFICIAL  
C
OGNITION
 
................................
................................
.........................
 
16
 
2.2
 
E
XAMPLES  OF  
C
OGNITI
VE  
A
RCHITECTURES
 
................................
................................
......................
 
18
 
2.2.1
 
Symbolic  Architectures
 
................................
................................
................................
..
 
18
 
2.2.2
 
ICARUS
 
................................
................................
................................
...........................
 
19
 
2.2.3
 
Prodigy
 
................................
................................
................................
..........................
 
20
 
2.3
 
E
MERGENT  
A
RCHITECTURES
 
................................
................................
................................
.........
 
21
 
2.3.1
 
NuPIC  
(Numenta  Platform  for  Intelligent  Computing)
 
................................
...................
 
21
 
2.3.2
 
NOMAD  
(Neurally  
Organised  Mobile  Adaptive  Device)
 
................................
.................
 
21
 
2.4
 
B
IOLOGICALLY  
I
NSPIRED  
C
OGNITIVE  
A
RCHITECTURES
 
................................
................................
........
 
22
 
2.5
 
H
YBRID  
A
RCHITECTURES
 
................................
................................
................................
..............
 
23
 
2.5.1
 
ACT
-­‐
R
 
................................
................................
................................
.............................
 
23
 
2.5.2
 
Kismet
 
................................
................................
................................
............................
 
24
 
2.5.3
 
CLARION
 
................................
................................
................................
........................
 
24
 
2.6
 
P
ROPOSED  
A
PPROACH  IN  
RUBICON
 
................................
................................
.............................
 
25
 
3
 
BASIS  OF  DESIGN:  INT
RODUCTION  TO  THE  SOF
NN
 
................................
......................
 
27
 
3.1
 
S
TRUCTURE  OF  THE  
SOFNN
 
................................
................................
................................
.........
 
28
 
3.2
 
P
ROGRESS  TO  DATE
 
................................
................................
................................
.....................
 
31
 
3.2.1
 
Application  of  the  SOFNN  for  function  approximation
 
................................
..................
 
32
 
3.2.2
 
Application  of  the  SOFNN  for  system  identification
 
................................
......................
 
33
 
3.2.3
 
Application  of  
the  SOFNN  for  time
-­‐
series  prediction
 
................................
.....................
 
35
 
4
 
REQUIREMENTS
 
................................
................................
................................
............
 
39
 
4.1
 
C
ASE  
S
TUDIES
 
................................
................................
................................
............................
 
39
 
4.2
 
I
NTERFACE  
R
EQUIREMENTS
:
 
L
EARNING  
L
AYER
-­‐
C
OGNITIVE  
L
AYER
 
................................
........................
 
47
 
4.3
 
I
NTERFACE  
R
EQUIREMENTS
:
 
C
OGNITIVE  
L
AYER
-­‐
 
C
ONTROL  
L
AYER
 
................................
........................
 
48
 
4.4
 
I
NTERFACE  
R
EQUIRE
MENTS
:
 
C
OGNITIVE  
L
AYER
-­‐
 
C
OMMUNICATION  
L
AYER
 
................................
............
 
49
 
4.5
 
N
ON
-­‐
FUNCTIONAL  AND  
C
OMMON  
R
EQUIREMENTS
 
................................
................................
..........
 
50
 
4.6
 
O
THER  
F
UNCTIONAL  
R
EQUIREMENTS
 
................................
................................
.............................
 
50
 
4.7
 
S
UMMARY
 
................................
................................
................................
................................
.
 
52
 
5
 
INTERFACE  SPECIFICAT
IONS
 
................................
................................
..........................
 
53
 
5.1
 
L
EARNING  AND  
C
OGNITIVE  
L
AYERS
 
................................
................................
................................
 
53
 
5.1.1
 
Data  structure  for  EVENTS
 
................................
................................
.............................
 
54
 
5.1.2
 
Map  between  OutputID  and  Event
-­‐
Location
 
................................
................................
.
 
55
 
5.1.3
 
Novelty  detection
 
................................
................................
................................
..........
 
56
 
5.2
 
C
OGNITIVE  AND  
C
ONTROL  
L
AYERS
 
................................
................................
................................
.
 
58
 
5.2.1
 
Proposed  control
-­‐
cognitive  goal  structure
 
................................
................................
....
 
58
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
6
 
 
5.2.2
 
Importance  of  the  ‘Parameters’  field
 
................................
................................
.............
 
59
 
5.2.3
 
Control  feedback  
status  structure
 
................................
................................
.................
 
61
 
6
 
HIGH
-­‐
LEVEL  DESIGN
 
................................
................................
................................
......
 
63
 
6.1
 
C
OGNITIVE  
M
EMORY
 
................................
................................
................................
..................
 
63
 
6.2
 
C
OGNITIVE  
R
EASONING
 
................................
................................
................................
...............
 
65
 
6.3
 
C
OGNITIVE  
D
ECISIONS
 
................................
................................
................................
.................
 
66
 
6.4
 
C
ONCLUSION
 
................................
................................
................................
.............................
 
67
 
7.
 
IMPLEMENTATION  PLAN  
FOR  NEXT  PERIOD
 
................................
................................
.
 
68
 
8.
 
ACKNOWLEDGEMENTS
 
................................
................................
................................
.
 
70
 
9.
 
REFERENCES
 
................................
................................
................................
..................
 
71
 
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
7
 
 
Abbreviations
 
ACT
-­‐
R  
 
 
Adaptive  Components  of  Thought
-­‐
Rational
 
BICA
 
 
Biologically
-­‐
Inspired  Cognitive  Architectures
 
CLARION  
 
Connectionist  Learning  Adaptive  Rule  Induction  ON
-­‐
line
 
DFNN  
 
 
Dynamic  Fuzzy  Neural  Network
 
EBF
 
 
E
llipsoidal  Basis  Function
 
EBL      
 
 
E
xplanation
-­‐
based  learning
 
FNN
 
 
Fuzzy  Neural  Network
 
GDFNN  
 
General
ise
d  Dynamic  Fuzzy  Neural  Network
 
IBCA  
 
 
Integrated  Biologically
-­‐
based  Cognitive  Architecture
 
MF
 
 
Membership  Function
 
MISO  
 
 
Multi  Input  Single  Output
 
MIMO
 
 
Multi  Input  Multi  Output
 
NL
-­‐
SOAR
 
Natural  Language  State,  Operator  and  Result
 
NOMAD  
 
Neurally  Organ
ise
d  Mobile  Adaptive  Device
 
NuPIC
 
 
Numenta  Platform  for  Intelligent  Computing
 
RBF
 
 
Radial  Basis  Function
 
RBFAFS  
 
Radial  Basis  Function  neural  network  based  Adaptive  Fuzzy  System
 
RMSE  
 
 
Root  
Mean  Squared  Error
 
RUBICON
 
Robotic  UBIquitous  COgnitive  Network
 
SOAR  
 
 
State,  Operator  
And
 
Result
 
SOFNN                        
Self
-­‐
Organizing  Fuzzy  Neural  Network
 
TS  
 
 
Takagi
-­‐
Sugeno  
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
8
 
 
W
SN
                             
 
Wireless  S
ensor  
N
etwork
 
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
9
 
 
Figures
 
F
IGURE  
1:
 
O
VERVIEW  OF  THE  COGNI
TIVE  LAYER  IN  
RUBICON
 
PROJECT
 
................................
...................
 
11
 
F
IGURE  
2
 
:
 
S
IMPLIFIED  HIGH  LEVEL
 
UNDERSTANDING  OF  THE
 
COGNITIVE  LAYER
 
................................
..........
 
13
 
F
IGURE  
3
 
:
 
C
ATEGORIES  OF  COGNITI
VE  ARCHITECTURES
[
11
]
 
................................
................................
.
 
18
 
F
IGURE  
4
 
:
 
T
HE  
SOAR
 
ARCHITECTURE  
[22]
 
................................
................................
.......................
 
19
 
F
IGURE  
5:
 
T
HE  
ICARUS
 
ARCHITECTURE
[18]
 
................................
................................
.....................
 
20
 
F
IGURE  
6:
 
P
RODIGY  COGNITIVE  ARC
HITECTURE
 
................................
................................
...................
 
21
 
F
IGURE  
7
 
:
 
A
 
REPRESENTATIVE  
ACT
-­‐
R
 
ARCHITECTURE  
[1,4]
 
................................
................................
.
 
23
 
F
IGURE  
8
 
:
 
T
HE  
CLARION
 
ARCHITECTURE
[40]
 
................................
................................
..................
 
25
 
F
IGURE  
9:
 
S
TRUCTURE  OF  THE  
SOFNN
 
................................
................................
............................
 
28
 
F
IGURE  
10
 
:
 
S
TRUCTURE  OF  THE  J
-­‐
TH  NEURON  
R
J
 
WITH  C
J
 
AND  Σ
J
 
IN  
EBF
 
LAYER
 
................................
.........
 
29
 
F
IGURE  
11
 
:
 
T
RAINING    RESULT  
(•
 
DESIRED  OUTPUT
,
 
+
 
NETWORK  OUTPUT
)
 
................................
..............
 
32
 
F
IGURE  
12
 
:
 
T
ESTING  RESULT  
(•
 
DESIRED  OUTPUT
,
 
+
 
NETWORK  OUTPUT
)
 
................................
................
 
33
 
F
IGURE  
13
 
:
 
T
RAINING  RESULT  
(

 
DESIRED  OUTPUT
,
 

 
NETWORK  OUTPUT
)
 
................................
...............
 
34
 
F
IGURE  
14
 
:
 
T
ESTING  RESULT
 
(

 
DESIRED  OUTPUT
,
 

 
NETWORK  OUTPUT
)
 
................................
................
 
34
 
F
IGURE  
15
 
:
 
T
RAINING  RESULT  
(

 
DESIRED  OUTPUT
,
 

 
NETWORK  OUTPUT
)
 
................................
...............
 
36
 
F
IGURE  
16
 
:
 
T
ESTING  RESULT  
(

 
DESIRED  OUTPUT
,
 

 
NETWORK  OUTPUT
)
 
................................
................
 
36
 
F
IGURE  
17:
 
T
HE  
MIMO
 
STRUCTURE  OF  THE  
SOFNN
 
................................
................................
.........
 
37
 
F
IGURE  
18:
 
H
IGH
-­‐
LEVEL  ARCHITECTURE  O
F  THE  COGNITIVE  LAYE
R
 
................................
..........................
 
63
 
F
IGURE  
19:
 
C
OGNITIVE  MEMORY  STRU
CTURE
 
................................
................................
....................
 
64
 
F
IGURE  
20:
 
C
OGNITIVE  REASONING  M
ODULE
 
................................
................................
.....................
 
65
 
F
IGURE  
21:
 
C
OGNITIVE  DECISIONS  M
ODULE
 
................................
................................
.......................
 
67
 
 
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
10
 
 
Tables
 
T
ABLE  
1
 
:
 
R
ESULTS  OF  FUNCTION  A
PPROXIMATION
 
................................
................................
..............
 
33
 
T
ABLE  
2
 
:
 
R
ESULTS  OF  SYSTEM  IDE
NTIFICATION
 
................................
................................
...................
 
35
 
T
ABLE  
3:
 
R
ESULTS  OF  TIME  SERIE
S  PREDICTION
 
................................
................................
...................
 
37
 
T
ABLE  
4:
 
 
-­‐
 
C
ASE  
S
TUDIES  IN  
AAL
 
S
CENARIO
 
................................
................................
.....................
 
39
 
T
ABLE  
5:
 
 
-­‐
 
C
ASE  
S
TUDIES  IN  
H
OSPITAL  
T
RANSPORT  
S
CENARI
O
 
................................
..............................
 
44
 
T
ABLE  
6:
 
 
-­‐
 
C
ASE  
S
TUDIES  FOR  BOTH  
S
CENARIOS
 
................................
................................
.................
 
46
 
T
ABLE  
7
 
:
 
T
HE  LEARNING  AND  COGN
ITIVE  LAYERS  INTERFA
CE  REQUIREMENTS
 
................................
............
 
47
 
T
ABLE  
8
 
:
 
T
HE  COGNITIVE  AND  CON
TROL  LAYERS  INTERFAC
E  REQUIREMENTS
 
................................
.............
 
48
 
T
ABLE  
9
 
:
 
I
NTERFACE  RE
QUIREMENTS  FROM  THE  
COMMUNICATION  LAYER
 
................................
................
 
49
 
T
ABLE  
10
 
:
 
N
ON
-­‐
FUNCTIONAL  AND  COMMO
N  REQUIREMENTS
 
................................
..............................
 
50
 
T
ABLE  
11
 
:
 
O
THER  FUNCTIONAL  
R
EQUIREMENTS
 
................................
................................
................
 
51
 
T
ABLE  
12
 
:
 
D
ATA  STRUCTURE  FOR  TH
E  LEARNING  AND  COGNI
TIVE  LAYER  INTERFACE
 
................................
...
 
54
 
T
ABLE  
13
 
:
 
M
AP  BETWEEN  
O
UTPUT
ID
 
AND  
E
VENT
-­‐
L
OCATION
 
................................
..............................
 
55
 
T
ABLE  
14:
 
P
ROPOSED  DATA  STRUCTU
RE  FOR  ISSUING  
G
OAL  
ID
S
 
................................
............................
 
58
 
T
ABLE  
15
 
:
 
G
OAL    
ID
 
EXPLOSION  FOR  SIMILA
R  ACTIVITIES
 
................................
................................
.....
 
60
 
T
ABLE  
16:
 
C
ONCISE  FORM  OF  THE  G
OAL  
ID
 
................................
................................
.......................
 
60
 
T
ABLE  
17:
 
D
ATA  STRUCTURE  FOR  CO
NTROL  STATUS  FEEDBAC
K
 
................................
..............................
 
61
 
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
11
 
 
1.

Introduction
 
1.1

Overview  of  the  
High  Level  
RUBICON  
Architecture
 
This  project  will  create  a  self
-­‐
sustaining,  self
-­‐
organizing,  learning  and  goal
-­‐
oriented  robotic  
ecology,  called  RUBICON  (Robotic  UBIquitous  COgnitive  Network),  where  a  robotic  ecology  is  
defined  as  a  network  of  heterogeneous  computational  nodes  interfaced  
with  sensors,  
effectors  and  mobile  robot  devices.  
 
 
The  nodes  of  
the  
RUBICON  ecology  mutually  support  one  another’s  learning.    RUBICON  
seeks  to  deliver  learning  solutions  yielding  
cheaper
,  
more  adaptive  and  more  efficient  
configuration  and  coordination  
of  
robotic  ecologies,  in  support  of  open,  dynamic,  
heterogeneous  and  computationally  constrained  systems,  as  well  as  a  wide  range  of  services  
and  end  user  applications.
 
 
Figure  
1
:  
Overview  of  
the  cognitive  layer
 
in  
RUBICON
 
project
 
 
Control  Layer

Cognitive  Layer

Learning  Layer

 
 
 
 
 
(Supervision)

Wiring,  Feedback,  Training

Execution  status

Goals

(Env  /  System)  

Events,  State

Sensor  data

Actions  (connamds)

Sensor  data

Messages

 
 
 
(Env  /  System)  

Events,  State,  confidence

 
(Novelty  Detection)

Feedback,Training

Sensing  and  
Actuation  (WSN/Robots/Human)

 
Action  
Quality

 
Communication

Layer

RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
12
 
 
The  RUBICON  architecture
 
consists  of  four  layers:  learning  layer,  control  layer,  cognitive  
layer  and  communication  layer.  A  high  level  architecture  of  the  RUBICON  project  is  shown  in  
Figure  
1
.    It  shows  interconnections  of  these  layers  highlighting  the  cognitive  layer.
 
 
The  overall  activities  of  these  layers  are  as  follows:
 


The  cognitive  layer  builds  up  knowledge  and  understanding  of  the  RUBICON  ecology  
based  on  reasoning  over  the  events  
learnt  by  the  learning  layer  and  information  on  
current  activities  (status)  received  from  the  control  layer
,  plus  experiential  
knowledge
.  It  facilitates  the  control  layer  with  the  decisions  on  ‘what  to  do’  under  
different  environmental  scenarios.
 


The  learn
ing  layer
 
provides  a  distributed,  adaptive,  and  self
-­‐
organizing  memory  
comprising  independent  learning  neurons  residing  on  multiple  nodes  of  the  RUBICON  
ecology.  These  neurons  interact  and  cooperate  through  the  underlying  
communication  channels  provided  by
 
the  Communication  Layer.
 


The
 
control  layer
 
provides  high  level  control  over  the  nodes  of  the  RUBICON
 
ecology,  
by  formulating  and  executing  both  action  and  configuration  strategies  to  sati
sfy  the  
objectives  set  for  the  
RUBICON  ecology  and  the  necessary  col
laborations  for  each  
node.  The  Control  Layer  uses  the  Learning  Layer  to  refine  the  
perception  capabilities  
of  the  
RUBICON  ecology  and  to  adapt  action  and  configuration  strategies  to  the  
environment.
 


The
 
communication  l
ayer
 
provides  inter
-­‐
component  communic
ation  and  integration  
mechanisms  by  leveraging  and  extending  state  of  the  art  solutions  in  WSNs  and  
middleware  for  robotic  ecologies.
 
Figure  
1
 
depicts
 
all  the  inter
-­‐
layer  dependencies  resulting  from  the  requirement
s
 
analysis  
and  specification  activities  
jointly  
presented  
in  
D1.1,  D2.1,  D3.
1
 
and  D4.1.
 
 
 
1.2

Role  of  the  Cognitive  Layer  in
 
the  RUBICON  project
 
The  overall  goal  of  this  work
-­‐
package  
(WP4)  
is  to  develop  cognitive  mechanisms  
that  enable  
the  
RUBICON
 
ecology  to  perform  reasoning  about  the  current  and  desired  states  of  the  
complete  
RUBICON
 
ecology,  and  in  addition  present  mechanism
s  that  permit  continuous  
exploration.
 
The  practical  objectives  are:
 


To  implement  an  advanced  cognitive  reasoning  system  within  the  
RUBICON
 
ecology
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
13
 
 


T
o  add  self
-­‐
adaptation  to  the  RUBICON  ecology  to  optimise  l
earning  and  information  
sharing.
 
 


To  implement  a  
novelty  detection  process,  whereby  events  of  particular  interest  are  
identified  and  utilised  as  motivational  drivers  for  exploration  and  active  learning,  thus  
constantly  enhancing  the  knowledge  embedded  in  the  RUBICON  ecology.  
 
The  
design  of  the  
cognitive  
layer  
is  based  on  
three  high  level  interconnected  modules.  They  
are  “cognitive  memory”,  “cognitive  reasoning”  and  “cognitive  decisions”.  
Figure  
2
 
depicts  
a  
high  level  layout  of  these  modules
 
and  their  interaction  with  the  learning  and  control  layers
.  
 
 
 
 
 
 
 

 

                 
 

 

 
 
 
 
The  cognitive  architecture  reflects  the  following:
 


The  “cognitive  memory”  module
 
stores  current  and  historical  information  f
rom
 
the  
ecology.  It  deals  with  the  inputs  to  the  cognitive  layer.  
 


The  
“cognitive  decisions”  module
 
is
 
responsible  for  
generating  
intermediate  
outputs  
from  the  cognitive  layer.  
Although,  t
he  final  output
s  come
 
from  the  cognitive  
decision  module
,
 
it  
exploits
 
output
s
 
from
 
the  cognitive  reasoning  
module  in  reaching  
 
 
 
 
 
 
 
 
 
 
Cognitive  layer
 
Cognitive
 
reasoning
 
Cognitive  decisions
 
Cognitive
 
memory
 
Learning  layer
 
Control  layer
 
Control  layer
 
Figure  
2
 
:  
Simplified  high  level  understanding  of  the  cognitive  layer
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
14
 
 
these  d
ecision
s
.  
The  o
utput
s
 
of  the  cognitive  reasoni
ng  
module  
are  
also  
u
sed  for  
internal  processing.
 


The  cognitive  layer
 
builds  a  high  level  understan
ding  from  information  on  discrete  
events
 
such  as  “door  is  open”,  “user  walked  out  of  the  bathroom”,  and  “faucet  is  
open”.
 


It  
also  
monitors  ongoing  activities  based  on  cognitive  scenario  assessments  
(“cognitive  reasoning”)  such  as  “user  cooking”,  “user  wate
ring  plant”.
 


It  decides  (“cognitive  decision”)  “goals”  for  the  
RUBICON
 
ecology,  such  as  “
room  1  
needs  to  be  cleaned”;  tap  in  bathroom  
needs  to  
be  
check
ed
”.  
 


It  provides  reasoning  on  
what
 
to  do  within  the  ecology  such  as  “check  status  of  the  
faucet  in  the  toilet”.  
 


The  control  layer  decides  
how
 
to  complete  the  requested  “goals”
 
and
 
sends  status  
updates  back  to  the  cognitive  layer.
 


The  cognitive  layer  builds  a
n
 
expanding  knowledge  base  of  t
he  ecology,  
based  on  its  
understanding  of  relevant  historical  
events,  
present  data
 
and  sensory  information
.
 
 
 
 
 
 
 
 
 
 
 
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
15
 
 
2

Overview  of  existing  Cognitive  Systems  Architectures  in  the  
Literature
 
Cognitive  architectures  are  frequently  used  to  model  human  behaviour  and  reproduce  such  
behaviour,  or  part  of  it,  in  artificial  systems.  Two  key  design  properties  of  cognitive  
architectures  are  
memory
 
and  
learning
.  Memory  holds  knowledge  about  the  world,  t
he  
agent  status  and  the  relationships  between  elements  of  the  world.  Learning  is  the  main  
process  that  shapes  memory.  These  two  elements  of  cognitive  architectures  underlie  the  
main  processes  that  are  considered  being  responsible  for  high
-­‐
level  functions  l
ike  reasoning,  
planning,  perception  and  inference  [
27
].  
Typically,  a
 
cognitive  system  builds  its  knowledge  
over  time  while  interacting  with  its  environment.  It  acquires  knowledge  through  
environmental  perception,  builds  knowledge  through  learning,  reasonin
g  and  planning  for  its  
actions  and  memorising  its  decisions  on  various  attended  events.  
 
Any  cognitive  architecture  is  designed  to  perform  certain  functions  based  on  its  core  
application  area.  These  function
s  include  [
based  on  
27
]
:
 


Recognition  and  
categorization
 
The  system  should  have  
the  
capabilit
y
 
to  recogn
ise
 
known  events  within  its  environment  
and  
be  able  to  
extract  features/patterns  from  those  events.    
 


Decision  making  and  choice
 
The  system  should  have  the  ability  to  produce  alternative  decisio
ns  
relevant  to  the  
impending  situation
.  Th
is
 
feature  make
s
 
a  cognitive  architecture  adaptable  to  various  
situations.  
 


Perception  and  situation  assessment
 
A  cognitive  architecture  may  be  connected  to  its  environment  through  various  sensory  
devices.  It  shoul
d  possess  the  capability  to  process  the  information  and  assess  the  
situation.
 


Prediction  and  monitoring
 
A  cognitive  architecture  builds  its  knowledge  over  a  period  of  time  by  monitoring  various  
states  of  its  components.  A  suitable  prediction  mechanism  can  
help  it  to  be  able  to  
handle  future  action  planning.  
 
 
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
16
 
 


Problem  solving  and  planning
 
The  system  should  be  able  to  generate  plans  to  complement  its  perceptions.  The  plan  
may  represent  a  set  of  actions  to  achieve  the  desired  goal.  It  is  obvious  that  all  of  t
he  
action  plans  may  not  succeed.  The  system  should  possess  a  problem  solving  capability  to  
deal  with  this  situation  by  utilizing  its  knowledge  and  resources.
 
Furthermore  the  
success/lack  of  success  should  be  recognised  and  utilised  in  the  overall  knowledge
 
base.
 


Reasoning  and  belief  maintenance
 
These  are  the  primary  requirements  of  any  cognitive  architecture  for  knowledge  
enhancement.  The  cognitive  architecture  should  relate  the  acquired  information  of  
events  with  its  existing  beliefs.  This  processing  helps
 
the  system  to  understand  the  states  
of  its  environment.
 


Execution  and  action
 
A  cognitive  architecture  should  be  able  to  represent  and  store  motor  skills  to  support  
and  drive  activities  in  the  environment.  Moreover  it  should  learn  new  skills  from  further  
i
nstructions  and  experience.
 


Interaction  and  communication
 
These  features  enable  a  cognitive  architecture  to  share  its  knowledge  with  other  systems  
in  the  environment.    
 


Remembering,  reflection,  and  learning
 
This  is  the  ability  to  encode  and  store  different
 
cognitive  processes  in  the  memory  for  
future  retrieval.  Moreover  the  architecture  should  broaden  its  knowledge  through  
learning,  hence  enhancing  its  beliefs.  
 
Although  a  selection  
of  
these  capabilities  are  common  to  many  of  the  proposed  cognitive  
architectures  in  the  literature  (particularly  learning  and  reasoning),  there  is  no  requirement  
for  
every  feature
 
to  appear  in  every  individual  architecture,  since  generally  
such  
architectures  
are  developed  for  a  specific  application  domain.  
 
 
2.1

Approaches  to  Artificial  Cognition
 
Among  the  various  approaches  to  cognition  we  can  discern  three  main  classes:
 


Symbolic  approach
 
based  on  symbolic  information  processing.
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
17
 
 


Emergent  approach  
based  on  connec
tionist  or  dynamical  systems,  where  self
-­‐
organisation  and  emergence  play  a  key  role  in  the  architecture.
 


Hybrid  approach  
based  on  a  combination  of  the  twos.
 
A
pproaches  base
d  on  biological  inspiration,  which  attempt  to  closely  mim
i
c
 
biology
,
 
are  
considered  
to  be  included  in  the  emergent  category.
 
Symbolic
 
architectures  focus  on  information  processing  using  high
-­‐
level  symbols  or  
declarative  knowledge.  This  is  usually  the  approach  of  standard  AI.  The  use  of  symbols  
supporting  information  processing  originates  
from  the  physical  symbol  systems  hypothesis  
[
27
]
,  which  states  that  symbol
 
manipulation  has  the  necessary  and  sufficient  means  for  
general  intelligence.  First  order  predicate  logic  is  usually  the  computational  approach  to  
symbols  manipulation.  Learning  is  
performed  by  asserting  predicates  in  a  knowledge  base.
 
SOAR  
(
State,  Operator  And  Result
)[
22
,  
23
]  and  
ICARUS
 
[
24
,  
25
]  
are  examples  of  this  type  of  
architecture
.
 
Emergent
 
architectures  are  composed  of  processing  nodes  connected  to  form  networks.  
Each  node  is  a  usually  simple  processing  element.  The  nodes  interact  with  each  other  
following  the  network  connections,  continuously  changing  their  internal  state.  The  overall  
beha
viour  of  the  architecture  is  therefore  the  emergent  result  of  the  behaviour  of  all  the  
single  nodes.  Neural  networks  are  a  common  example  of  emergent  architectures.  The  
mechanisms  behind  learning  are  highly  dependent  on  the  network  model  in  use.  A  common  
w
ay  to  store  knowledge  is  to  change  the  weights  of  the  links  connecting  the  nodes,  either  in  
a  global  way  (e.g.  the  perceptron  rule)  or  in  a  local  way  (e.g.  Hebbian  learning).
 
IBCA  
(
Integrated  Biologically
-­‐
based  Cognitive  Architecture
)  [
38
]  and  
NOMAD  
(Neurally  Organ
ise
d  
Mobile  Adaptive  Device)  [
12
]  are  well  known  examples.
 
Hybrid  architectures  are  combinations  of  the  above.  Symbolic  architectures  are  able  to
 
process  high
-­‐
level  information  in  a  way  that  resembles  human  expertise.  However  they  are  
not  su
itable  for  processing  raw  data,  such  as  sensor  streams  or  images.  Conversely,  
emergent  architectures  are  better  suited  to  handle  large  amount  of  data  and  uncertainty,  
and  to  generalise  to  unforeseen  situations.  Yet  they  lack  the  capability  to  real
ise
 
high
-­‐
level  
cognitive  functions.  Therefore  attempts  have  been  made  to  combine  symbolic  manipulation  
and  connectionism  in  hybrid  architectures.
 
ACT
-­‐
R  
(
Adaptive  Components  of  Thought
-­‐
Rational
)  [
2,  3
]  and  
CLARION  
(The  
Connectionist  Learning  Adaptive  Rule  Induction  
ON
-­‐
line
)  
[
40
]  are  examples  of  hybrid  network  architectures.
 
Figure  
3
   
summarises  the  overall  properties  of  the  Symbolic,  Emergent  and  Hybrid  cognitive  
architectures  (Duch  et  al.  [
11
])
.
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
18
 
 
 
Figure  
3
 
:  Categories  of  cognitive  architectures[
11
]
 
 
2.2

Examples  of  Cognitive  Architectures
 
Below  is  a  s
hort  survey  of  some  of  the  best
-­‐
known  architectures  that  are  still  being  
developed.  More  in
-­‐
depth  surveys  may  be  found  in  [
11
]  and  [
23
].
 
2.2.1

Symbolic  Architectures
 
SO
AR  
(State,  Operator  and  Result)
 
is  a  rule
-­‐
based  cognitive  architecture  designed  to  model  
general  knowledge  [
22
].  SOAR  stores  the  knowledge  in  
the  form  of  production  rules
 
arranged  in  terms  of  operators  that  act  on  symbols  and  predicates.  It  operates  by  alternating  
production  with  decision.  Production  rules  are  activated  depending  on  the  current  goal  or  on  
the  status  of  the  declarative  memory.  This
 
activation  modifies  the  m
emory,  
which  in  turn  
triggers  other  production  rules.  When  no  more  production  rules  are  activated,  the  decision  
cycle  begins.  Decision  is  performed  by  activating  sub
-­‐
goals  according  to  action  preferences.  
This  in  turn  might  change  the  memory
,
 
which  will  tr
igger  a  new  cycle  of  production  rules  
activation.  
An  overview  of  the  architecture  is  shown  in  Figure  4.
 
The  learning  mechanism  in  SOAR  is  called  chunking  which  is  an  EBL  (
explanation
-­‐
based  
learning
)  technique  for  formulating  rules  [
22
].  The  decision  proced
ure  selects  
operators  
and  
detects  
impasses.  It  
matches  productions  against  elements  in  the  working  memory  to  
generate  sub
-­‐
goals  automatically  when  it  detects  impasse.  If  a  sub
-­‐
goal  overcomes  the  
problem  then  SO
AR  adds  a  new  chunk  to  the  long
-­‐
term  memory.  S
ub
-­‐
goals  can  also  
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
19
 
 
deliberately  access  episodic  or  semantic  memory  to  retrieve  knowledge  relevant  to  resolving  
the  impasse.  This  process  leads  to  the  dynamic  generation  of  a  goal  hierarchy.  Researchers  
have  used  SOAR  to  develop  a  variety  of  sophisticated  ag
ents  and  high
-­‐
level  cognitive  
functions  that  have  demonstrated  impressive  functionality  such  as  computer  games,  human  
language  processing
 
and  
categorization  [
22
,  
29
,  
31
,  
35
,  
44
].
 
SOAR  has  been  applied  to  high
-­‐
level  planning  problems  and  natural  language  
comprehension  (NL
-­‐
SOAR).
 
 

Figure  
4
 
:  The  SOAR  architecture
 
[
22
]
 
2.2.2

ICARUS
 
ICARUS  
is  a  cognitive  architecture  designed  for  physical  agents
 
[
24
]
.  ICARUS  relies  on  
symbolic  manipulation  of  knowledge,  but  it  distinguishes  between  concepts  and  skills,  each  
of  them  having  a  different  memory  representation.  The  architecture  includes  modules  for  
perception,  planning  and  execution.  Concepts  are  processed
 
in  a  bottom
-­‐
up  way  to  match  
percepts,  while  goals  are  processed  in  a  top
-­‐
down  way  to  match  skills.  
 
Figure  
5
 
depicts  the  ICARUS  architecture.  It  
store
s  two  distinct  forms  of  knowledge  known  as  
concepts  and  skills.  The  concepts  deal  with  environmental  situations  in  terms  of  other  
concepts  and  percepts  and  skills  deals  with  ordered  sub
-­‐
goals  to  achieve  goals  [
26
].  
Moreover,  skills  refer  to  initiation  and  
continuation  conditions  of  concepts.  The  
ICARUS  
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
20
 
 
architecture  achieves  its  knowledge  through  hierarchical,  incremental  reinforcement  
learning  and  propagates  reward  values  backward  through  time  [
11
].  It  creates  
a  new  skill  
with  one  or  more  ordered  sub
-­‐
goals  
whenever  its  problem  solver  achieves  a  new  goal.  It  
uses  that  knowledge  and  executes  the  skill  if  the  system  encounters  similar  problems  in  the  
future  [
24
].  Researchers  have  used  
ICARUS  
to  develop  agents  for  a  number  of  domains  that  
involve  a  combination  o
f  inference,  execution,  problem  solving,  and  learning  [
9
,  
19
,  
25
].  
 
 
 
Figure  
5
:
 
The  ICARUS  architecture
[18]
 
2.2.3

Prodigy
 
Prodigy,  shown  in  
Figure  
6
,  is  a  cognitive  architecture  in  which  the  learning  and  reasoning  
modules  produce  mutually  interpretable  knowledge  structures  [
5
].  Its  general
-­‐
purpose  
problem  solver  searches  through  a  problem
 
space  to  accomplish  a  set  of  goals  from  a  
specified  initial  state  description.  This  search  relies  on  a  set  of  
control  rules,
 
which  
may  be  
general  or  domain  specific,  hand
-­‐
coded  or  automatically  acquired,  and  may  consist  of  
heuristic  preferences  or  definit
ive  selections.
 
The  control  rules  are  based  on  ‘select’,  ‘reject’  
and  ‘prefer’.  
On  each  cycle,  
Prodigy  
uses  its  control  rules  to  select  an  operator,  binding  set,  
state,  or  goal.  If  unsuccessful,  it  assigns  the  whole  set  of  candidates  and  applies  the  reject
ion  
rules  to  filter  them  out  gradually  and  finally  the  most  preferred  alternative  is  achieved  by  
using  the  preference  rules.  Various  aspects  of  the  Prodigy  architecture  can  be  found  in  [
6,  
15
].  
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
21
 
 
 
Figure  
6
:  Prodigy  cognitive  
architecture
 
 
2.3

Emergent
 
Architectures
 
2.3.1

NuPIC  
(Numenta  Platform  for  Intelligent  Computing)  
 
NuPIC
 
is  a  connectionist  architecture  based  on  hierarchical  temporal  memory.  The  nodes  are  
organised  in  a  hierarchical  way,  each  of  them  implementing  learning  and  
memorisation.  
NuPIC's  main  feature  is  the  support  for  dynamical  pattern  learning,  allowing  information  
that  unfolds  with  time  to  be  stored  and  processed.  Pattern  anticipation  is  therefore  naturally  
occurring  in  this  architecture.
 
2.3.2

NOMAD  
(Neurally  Organ
ise
d  
Mobile  Adaptive  Device)
 
NOMAD  is  
also  known  as  Darwin  automata  [
12
],  is  mainly  used  for  real
-­‐
time  pattern  
recognition.  A  reward  mechanism  drives  learning  in  a  large
-­‐
scale  neural  network.  Spatial  and  
episodic  memory  is  obtained  by  a  modelled  hippocampus.  It
 
has  been  used  for
 
visual  scene  
understanding  and
 
invariant  object  recognition.
 
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
22
 
 
2.4

Biologically  Inspired  Cognitive  Architectures
 
A  number  of  cognitive  architectures,  which  draw  more  specific  motivations  from  the  
biological  brain  have  also  been  proposed  and  a
re  in  various  stages  of  development.    These  
include  
the  Novamente  AI  Engine  
[
30
],  which  relates  principles  of  neurodynamics  and
 
associated  emergent  patterns.  In  this  approach,  self
-­‐
organizing  and  goal
-­‐
oriented  
interactions  between  patterns  are  responsible  for  mental  states.  Hierarchical  and  relational  
pattern  organsiation  emerges  from  the  dynamics  of  network  activations.  A  tree
-­‐
lik
e  
structure  models  actions  and  perceptions.  The  
4CAPS  
architecture  [
20
]  has  been  designed  to  
have  a  biologically  plausible  neural  architecture.  It  assumes  that  cognition  arises  from  the  
concurrent  activation  of  multiple  nodes  in  a  collaborative  manner.  Eac
h  centre  corresponds  
to  a  specific  brain  region,  but  the  topology  of  the  complete  network  is  adaptable,  and  each  
of  these  exhibits  limited  computational  resource.  The  
AMBR
 
cognitive  systems  attempts  to  
implement  a  model  of  human  reasoning,  perception,  memo
ry,  human  judgement  and  
deduction.  The  DARPA  Biologically
-­‐
Inspired  Cognitive  Architectures  (BICA)  program  has  
proved  very  fruitful,  spawning  the  BICA  conference  and  the  TOSCA  proposal  (TOSCA  
Comprehensive  brain
-­‐
based  model  of  human  mind”  [
46
]).    
Tahboub  et
 
al  [
42
]  have  
presented  a  neuro
-­‐
fuzzy  reasoning  system  for  mobile  robotics,  exploiting  the  synergy  
between  neural  system  learning  capabilities  in  parallel  with  human
-­‐
like  fuzzy  reasoning.  
 
Homma  and  Gupta  consider  a  
Fuzzy  Self
-­‐
organizing  Map  as  an  emulatio
n  of  the  Cerebral  
Cortical  Structure  for  Pattern  Recognition  
[
17
]
 
and
 
demonstrate  a  neural  structure  for  
formation  of  long
-­‐
term  memory.  Their  approach  takes  feedback  from  cognitive  results  based  
on  the  current  long
-­‐
term  memories,  repre
senting  the  current  knowledge,  
and  suggest  that  
the  approach  emulates  biologically  observed  features  of  human  memory.  Gupta  et  al  also  
consider  computational  perception  and  cognition  under  
uncertainty  
in  [
14
],
 
noting  the  great  
tolerance  in  human  cognition  
for  imprecision  or  uncertainty,  and  the  advantages  of  fuzzy  
systems  in  model
l
ing  the  latter.
 
A  very  interesting  approach  is  that  of  
Erlhagen  and  Bicho  [
13
]  who  proposed  a  dynamic  
neural  field  approach  to  cognitive  robotics,  inspired  by  current  understandin
gs  of
 
the  
processing  principles  and  the  neuronal  circuitry  underlying  memory,  decision  making,  action  
understanding  and  prediction  in  the  biological  brain.  The  approach  uses  a  coupled  system  of  
dynamic  neural  fields,  each  representing  the  basic  functionality  of  
neuronal  populations  in  
different  brain  areas.  
 
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
23
 
 
2.5

Hybrid  Architectures
 
2.5.1

ACT
-­‐
R
 
ACT
-­‐
R  
(
Adaptive  Components  of  Thought
-­‐
Rational
)  is  a  hybrid  cognitive  architecture  that  
has  
the  primary  goal  of  mode
l
ling  human  behaviour
 
[
2
]
.    It  seeks  to  b
uild  a  system  that  can  
perform  
a
 
full  range  of  human  cognitive  tasks  and  describe  in  detail  the  mechanisms  
underlying  perception,  thinking,  and  action  [
22
]
.  ACT
-­‐
R  uses  a  top
-­‐
down  learning  approach,  
where  symbolic  constructs  (chunks)  are  created  to  describe  the  result  of  a  complex  
operation,  and  subsequently  matched  by  production  rules  operated  by  the  sub
-­‐
modules.  
 
Its  
symbolic
 
structure  represents  a  producti
on  system,  whereas  the  
sub
-­‐
symbolic
 
structure  is  
represented  by  a  set  of  massive  parallel  processes  represented  by  a  number  of  
mathematical  equations.  The  sub
-­‐
symbolic  equations  control  many  of  the  symbolic  
processes  [
1
].  ACT
-­‐
R  has  three  main  components:  m
odules,  buffers  and  pattern  matcher.  
The  architecture  is  composed  of  intentional,  declarative,  visual,  manual  and  goal  modules.  
There  are  two  types  of  modules  named  perceptual
-­‐
motor  modules  and  cognitive  memory  
modules  (
Figure  
7
 
[
4
]).  
 
 
Figure  
7
 
:  A  representative  ACT
-­‐
R  architecture
 
[1,4]
 
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
24
 
 
P
erceptual
-­‐
motor  modules  create  
interfaces  with  the  real  world.  There  are  two  types  of  
memory  modules  named  
declarative  memory  (
consists  of  facts)  and  
procedural  memory  
(
represent  knowledge  about  different  procedures/productions).  
A  production  system  
coordinates  the  action  of  the  modules
 
by  matching  the  information  stored  in  several  buffers,  
where  each  module  places  a  limited  amount  of  information.  
ACT
-­‐
R  accesses  its  modules  
through  interfaces  character
ise
d  by  dedicated  buffers  whose  contents  represent  the  state  of  
ACT
-­‐
R.  The  pattern  matc
her  searches  for  a  production  that  matches  the  current  state  of  the  
buffers.  The  state  of  the  system  can  be  changed  by  execution  of  the  chosen  production.  
The  
ACT
-­‐
R  architecture  is  used  in  different  application  areas  including  different  aspects  of  
memory,  
perception,  attention,  reasoning  and  decision  making,  problem  solving,
 
and  
language  processing  [
21
,  
41
,  
45
].
 
2.5.2

Kismet
 
Kismet  
is  the  cognitive  architecture  underlying  the  homonymous  articulated  
ant
h
ropomorphic  robotic  head  [
26
].  Kismet  engages  people  in  
expressive  face
-­‐
to
-­‐
face  
dialogues,  trying  to  understand  the  dynamics  of  social  interaction  by  cues  like  gaze  direction,  
facial  expression  and  vocal  babbling.  Kismet  has  five  distinct  modules  in  its  cognitive  
architecture:  a  perceptual  system,  an  emotion  sy
stem,  a  behavio
u
r  system,  a  drive  system,  
and  a  motor  system.  
 
The  robot  behaviour  and  the  robot  response  to  the  environment  
are
 
modulated  by  
emotions.  They  motivate  the  system  and  enable  the
 
learning  of  new  behaviours.  
Another
 
type  of  motivation  in  Kismet
 
is
 
drive
s
,  which  establish  the  top
-­‐
level  goals  of  the  robot.  Drives  
like  

rest

 
or  

engage  toys

 
focus  the  robot  behaviour  using  cognitivist  rule
-­‐
based  schema  to  
pursue  the  goals,  but  the  overall  robot's  behaviour  emerges  from  the  interaction  of  all  the  
s
ub
-­‐
modules.
 
2.5.3

CLARION
 
CLARION  
is  a  hybrid  architecture  that  
stores  both  action
-­‐
cen
tred  and  non
-­‐
action  cent
red  
knowledge  in  implicit  form  using  multi
-­‐
layer  neural  networks  and  in  explicit  form  using  
symbolic  production  rules  [
27
,  
40
].  
Figure  
8
   
shows  an  overview  of  the  CLARION  architecture  
[
40
];  i
t  
captures  the  interactions  between  the  e
xplicit  and  implicit  knowledge
.    
Sensory  
infor
mation  is  passed  to  the  implicit  layer.  It  generates  alternative  high
-­‐
value  actions.  The  
explicit  layer  uses  rules  to  propose  actions.  
CLARION  also  employs  different  learning  
methods  such  as  reinforcement  learning  methods  or  back
-­‐
propagation  for  weight  rev
ision,  
estimation  of  value  functions  and  construction  of  production  rules.
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
25
 
 
 
Figure  
8
 
:  The  CLARION  architecture
[40]
 
 
2.6

Proposed  Approach  in  
RUBICON
 
Our  analysis  of  approaches  to  cognitive  architectures  is  dictated  by  a  number  of  conclusions  
and  observations,  which  emerge  from  the  cognitive  systems  literature,  as  applied  to  
robotics,  coupled  with  considerations  of  pragmatic  limitations  specific  to  RUBI
CON.    
 
Primary  among  these  is  the  observation  that  while  s
ymbolic  AI  approaches  have  achieved  
considerable  success  in  specific  areas  (language,  pattern  
recognition,  memorization  and  
retrieval  of  vast  amount  of  information  etc)  and  have  
dominated  the  tradit
ional  AI  approach,
 
it  is  not  yet  clear  whether  these  approaches  will  lead  to  
effective  
higher  level  cognition.
 
Furthermore,  it  is  clear  that  r
esearch  in  autonomous  robotics  has  illustrated  that  reasonably  
complex  robot  behavio
u
rs  may  be  organ
ise
d  without  the  need  for  high
-­‐
level  symbolic  
representations.  
 
The  Rubicon  ecology  also  has  several  resource
-­‐
constrained  attributes,  which  will  
substantially  limit  the  application  of  a  complex  and  substantial  high
-­‐
level  symbolic  approach.  
Furthermore,  the  hi
gh  level  architecture  of  Rubicon  has  largely  segmented  the  learning  and  
reasoning  components  of  the  overall  system.  
 
A  tight  coupling  between  Events  (learn
t
 
by  the  learning  layer)  and  Goals  (set  for  the  Control  
layer)  is  a  requirement  of  the  Rubicon  ecolog
y.  The  traditional  closed  loop  between  sensors  
and  effectors  enables  a  fast  real
-­‐
time  adaptation  to  dynamic  scenarios,  but  fails  to  address  
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
26
 
 
somewhat  higher  cognitive  functions  which  relay  on  working  (episodic)  memory,  past  
experiences,  and  a  perception  of  
a  desired  and  preferred  world  order  which  cannot  be  solely  
related  to  a  series  of  current  sensory  inputs.  There  is  a  requirement  not  only  to  address  the  
implications  of  the  current  series  of  
incoming
 
multimodal  sensory  data,  but  to  mediate  the  
implicated  g
oals  with  both  a  historical  memory  of  past  decisions  and  events  and  a  broader,  
more  global  consideration  of  the  overall  ecology’s  status.  This  may  for  example  imply  the  
ability  to  interpret  the  current  sensory  data  stream  in  the  context  of  an  ongoing  lon
g
e
r
-­‐
term  
global  goal  or  set  of  goals,  or  a  specific  trend  towards  a  particular  disastrous  condition.  
Further
more  it  is  recogn
ise
d  that
 
the  portfolio  of  events  and  portfolio  of  possible  goals  is  
dynamic,  requiring  a  self
-­‐
organising  system.
 
Our  conclusion  ther
efore  is  that  for  the  Cognitive  Layer  in  Rubicon  we  should:
 


Avoid  a  computational
ly
 
demanding  symbolic  based  approach
 
if  at  all  possible
 


Apply  an  approach  that  recognises  the  existence  of  a  distinct  and  separate  learning  layer
 
in  the  ecology,  while  working
 
synergistically  with  this  module
 


Exploit  the  potential  of  a  biologically  inspired  architecture  that  mimics  aspects  of  the  
biological  brain  and  offers  both  compatibility  with  the  spiking  neural  network  based  
reservoir  computing  approach  of  the  Learning  Lay
er,  but  also  offers  potential  for  future  
proofing  the  architecture  by  progressively  incorporating  more  and  more  emulations  
(albeit  probably  first  order  models)  of  biological  brain  functions  and  regions  of  interest
 


Place  self
-­‐
organisation  as  a  core  design  c
oncept.
 
We  therefore  approach  the  problem  of  an  appropriate  cognitive  architecture  for  Rubicon  
from  a  more  computational  intelligence  
(as  opposed  to  general  artificial  intelligence)  
angle,  
whereby  we  seek  to  learn  from  the  signal  processing  methodology  evi
dent  in  the  biological  
brain.  Neural  based  approaches,  dynamic  neural  fields,  fuzzy  human
-­‐
like  reasoning  and  self
-­‐
organising  neuro
-­‐
fuzzy  techniques  are  therefore  central  to  the  cognitive  layer  design  whose  
initial  concepts  are  illustrated  later  in  this  doc
ument.
 
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
27
 
 
3

Basis  of  Design:  Introduction  to  the  SOFNN
 
It  is  now  widely  recognised  that  fuzzy  logic  offers  a  very  powerful  framework  for  approximate  
reasoning  as  it  attempts  to  model  the  human  reasoning  process  at  a  cognitive  level.    Similarly,  
neural  networks  offer  a  highly  structured  architecture,  with  learn
ing  and  generalisation  
capabilities,  which  attempts  to  mimic  the  neurological  mechanisms  of  the  brain.    Their  fusion  
in  fuzzy  neural  network  architectures  provides  a  powerful  framework  for  cognition  and  in  
particular  the  reasoning  and  decision  functionalit
y  outlined  in  
S
ection
 
2
.  
 
This  section  will  review  the  theoretical  basis  for  the  design  of  a  Self
-­‐
Organizing  Fuzzy  Neural  
Network  (SOFNN).  The  aim  is  to  create  a  new  SOFNN  that  reflects  the  knowledge  being  
obtained  by  wireless  sensor  nodes  (WSN)  with  an  ad
aptable  network  configuration.  
 
The  SOFNN  is  essentially  a  mathematical  concept  without  physical  embodiment;  in  contrast,  
the  RUBICON  will  have  both  physical  embodiment  and  geographical  location.    Therefore,  this  
work
-­‐
package  will  review  the  theoretical  ba
sis  underpinning  the  SOFNN  approach,  and  
extend  it  to  permit  incorporation  of  position  based  sensor  location  information  and  
attributes.  
 
In  the  SOFNN,  neurons  are  added  or  pruned  based  on  specific  performance  criteria,  using  a  
structured  learning  approach
.  To  do  this,  we  will  represent  each  node  in  the  ecology  as  
either  a  single  or  group  of  ellipsoidal  basis  function  (EBF)  neurons  with  a  centre  vector  and  a  
width  vector.  The  EBF  neuron  represents  a  fuzzy  rule  formed  by  AND  logic  (or  T
-­‐
norm)  
operating  on  Ga
ussian  fuzzy  membership  functions.  The  elements  of  centre  vector  and  width  
vector  of  the  EBF  neuron  are  the  centres  and  widths  of  the  Gaussian  membership  functions.  
The  challenge  will  be  to  incorporate  position  based  information  into  the  structured  learnin
g  
algorithm  so  as  to  ensure  the  integrated  RUBICON  environment  and  that  the  dynamic  
restructuring  will  support  continuous  learning.
 
The  core  challenges  and  requirements  of  this  work
-­‐
package  are  therefore  to  create  a  
framework  whereby  the  overall  system  dec
ides  autonomously:
 


how  to  develop  a  SOFNN  which  accurately  reflects  the  dynamics  of  the  ecology  in  
both  physical  embodiment  and  cognitive  capability;
 


devise  suitable  strategies  for  neuron  addition,  as  well  as  pruning  of  unnecessary  or  
unimportant  nodes;
 


wh
en  and  how  to  explore  an  unusual  event  to  support  learning  driven  by  novelty  so  
as  to  drive  continuous  self
-­‐
adaptation  and  self
-­‐
organisation.
 
For  learning  a  fuzzy  rule
-­‐
based  model,  there  are  two  parts:  structure  learning
 
and  parameter  
learning.  Structure  l
earning  tries  to  find  how  many  rules  and  membership  functions  are  
RUBICON  D4.1  
First  Synopsis  of  the  RUBICON  Self  Organising  
Fuzzy  Neural  Network  (SOFNN)
 
 
 
RUBICON
:  Project  No.:  269914
 
 
3
0/9
/2011
 
Page  
28
 
 
necessary  to  model  the  available  data.  Parameter  learning  is  concerned  with  the  decisions  
about  the  parameters  of  membership  functions  in  the  premise  part  and  the  linear  crisp  
function  in  th
e  consequence  part.  For  neural  networks  based  on  fuzzy  inference  systems  and  
self
-­‐
organization  fuzzy  logic  systems,  the  main  limitation  is  that  they  cannot  automatically  
acquire  the  fuzzy  rules  that  they  use  to  make  their  decisions.  Neural  networks  have  th
e  
ability  to  learn  the  fuzzy  rules  automatically.  But  usually  it  is  not  possible  to  extract  the  fuzzy  
rules  from  the  trained  neural  networks  [
36
].  Fuzzy  neural  networks  (FNNs)  are  hybrid  
systems  that  combine  the  theories  of  fuzzy  logic  and  neural  networks.
 
In  these  hybrid  
systems,  the  fuzzy  techniques  are  actually  used  to  create  or  enhance  neural  networks  [
37
]  
and  can  be  used  to  learn  membership  functions  and  create  fuzzy  rules  that  may  be  easily  
interpreted.
 
 
3.1

Structure  of  the  SOFNN
 
The  self
-­‐
organizing  fuzz
y  neural  network  (SOFNN)  [
28
]  is  a  five
-­‐
layer  network  as  shown  in  
Figure  
9
.  The  five  layers  are  the  input  layer,  the  ellipsoidal  basis  function  (EBF)  layer,  the  
normalised  layer,  the  weighted  layer,  and  the  output  layer.  Some  of  the  interconnections  are