Download Modeling context-aware distributed knowledge

AAAI Spring Symposium 2003 on
Agent-Mediated Knowledge Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Modeling context-aware distributed
knowledge
Jorge Louçã
[email protected]
ISCTE – Instituto Superior de Ciêncas do Trabalho e da Empresa
Lisboa, Portugal
1
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Main Goal and Proposition
Research goal :
to model multi-dimensional decision-making processes using artificial
agents
Propositions :
• use causal cognitive maps to model software agents
• compose a collective solution to a goal through a distributed and
incremental process, based on interactions
• use context in cognitive maps to define agent’s mental states
2
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Philosophy and artificial intelligence
Theoretical foundations: AI as a laboratory
Philosophy and Artificial Intelligence: both try to understand all different kinds
of perception, action and intelligence
The association of this domains allows :
• to simulate reasoning in software programs, starting from a given
conception of what can be the mind;
• to do controlled experiments aiming to understand the knowledge
representation systems used in our mind to represent the world;
AI as a laboratory :
• a specified idea of what is the mind leads to experiments about new
software architectures
• experiments can be seen as a way of really doing philosophy, because
they search the conditions that make possible cognition in general - human
intelligence
3
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Philosophy and artificial intelligence
Theoretical foundations: AI as a laboratory
Intelligence:
the capacity of problem solving and decision
There is no problem solving and decision capacity without some
representation of the world
Folk Psychological Constructs help us to model the knowledge representation
systems used by our mind, using mental states
Advantages of Folk Psychology : operational, comprehensible, an instrument
to predict and explain behavior, to manipulate mental states
4
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Philosophy and artificial intelligence
Theoretical foundations: Functionalism
This use of mental states can act as a functionalist theory if we identify
mental states in terms of their causal-functional relationships
Functionalism identifies mental content with causal-functional roles
[Goldman,93]
Causal Covariance Theory of Content [Allen,02]
proposes that mental states get their content by being causally related
only to what they are about (e.g., to those mental states belonging to its
own specific context)
This idea of using context to define mental states is adopted to manipulate
mental states in a Multi-Agent System, through the use of Cognitive Maps
5
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Cognitive Mapping
Theoretical foundations:
using cognitive maps to model individual and collective beliefs
Cognitive Map : graphical representation of individual or collective beliefs,
regarding a specific domain
Cognitive mapping is used by psychologists and decision
makers to :
understand the behavior of actors participating in a decision making process
detect conflicts (incoherent viewpoints)
discuss points of view
++
State of
the wold
Task
Goal
very positive
Influence
fortementinfluence
positive
+
-
Influence
positive
positive
influence
Influence negative
--
negative influence
Influence fortement negative
[Carlson et Walden,96]
very negative influence
6
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Cognitive Mapping
Theoretical foundations:
using cognitive maps to model individual and collective beliefs
t6 - Create
propositions
+
e1 - Age of
employees :
30 / 35
+
(-,++)
e2 - Employees
notion of
"group"
t1 - Interact
(- -,++)
(+,++)
t2 - Internal
debates
b2 - Inovate
working
processes
++
++
t8 - Invest in R & D
+
t9 - Professionnal
learning
+
++
t7 - Research to
improve working
processes
e3 - Employees
confidence
b1 - Adapt
employees to
changes
(-,+)
--
t3 - Accept
suggestions
+
t4 - Periodical
reunions
+
+
t5 - Dialogue
e4 - Resistance to
change working
habitudes
Example of a partial cognitive map [Louçã,00]
Reflexive character of CM - we each construct our private versions of reality and deal
only with those constructions, which may or may not correspond to some real world
[Lissack & Ross,99]
7
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Cognitive Mapping
Theoretical foundations:
using cognitive maps to model individual and collective beliefs
Collective interpretation using CM - according
to Karl Weick and others, organizations can also
be seen, at another abstraction level, as systems
of construction and interpretation of reality
[Weick,95] [Lissack & Gunz,99]
e1 - Age du
e2 - Cohésion
personnel :
30 / 35
+
t1 - Interagir
construction
t2 - Faire
réflexion
interne
and
+
t9 - Faire formati
professionnelle
+
t3 - Accepter
suggestions
b1 - Adapter le
personnel au
changement
+
---
+
+
t4 - Faire des
réunions
périodiques
du personnel
++
++
Different
levels
of
interpretation of reality:
individual
collective
e3 - Confiance
du personnel
-
t5 - Dialoguer
e4 - Résistances à
changer d'habitudes
de travail
CM facilitate reasoning, communication and discussion about individual and
corporate knowledge
8
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Research in AI using Cognitive Maps
Theoretical foundations: state of the art
[ M a r c h a n t,
99]
[ W e llm a n ,
94a]
[ C o n s o le e t
a l.,8 9 ]
[Z h an g et
a l.,
92 et 94]
[P a rk &
K im ,
95]
IA d o m a in
In f e re n c e
fro m fu z z y
im p lic a tio n
In fe r e n c e in
cau sal
n e tw o r k s
C o g n itiv e
m ap as a
lo g ic a l
s y s te m
C ausal
in f e r e n c e ;
MAS ;
n e u ra l
n e tw o rk s
N P N L o g ic
C ausal
in f e r e n c e
w ith c ir c u its
F o rm a l
m odel
C au sal
in f e r e n c e in
q u a lita tiv e
p r o b a b ilis tic
n e tw o r k s
C o g n itiv e
M ap
K in d o f
lin k s
b e tw e e n
c o n c e p ts
R e a s o n in g
fro m
C o g n itiv e
M aps
D e te c t a n d
s o lv e
c o n flic ts in
MAS
Fuzzy
im p lic a tio n
P r o b a b ilis tic
c a u s a lity
C a u s a lity
yes
yes
no
no
C a u s a l lin k
as an
im p lic a tio n
f o r m u la
[ C h a ib d ra a ,
97a, 97b et
98]
C ausal
in fe r e n c e ;
MAS
C o g n itiv e
M ap
C a u s a lity
L o g iq u e e t
R e la tio n s
N P N L o g ic
and
R e la tio n s
w ith c ir c u its
C a u s a lity
R eason
f ro m c a u s a l
n e tw o r k s
yes
yes
yes
no
yes
no
yes
C a u s a lity
9
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Modeling context-aware distributed knowledge
Proposition: distributed and incremental reasoning process
An agent uses the set of concepts
represented in its cognitive map. The
agent composes its individual solution
to the goal, represented by a partial
cognitive map.
Each agent that receives an allocation
message including a sub-goal, starts
its own reasoning process and, in
return, responds with a solution to the
sub-goal. This distributed reasoning
process allows representing several
points of view concerning the subgoals.
Individual solution
to a goal
Goal allocation : results from other reasoning processes
Solution
to an allocated goal
Solution
to an allocated goal
Solution
to an allocated goal
----------------Cognitive Map Aggregation-------------------------
Collective Solution
Mat
rix
c
d en
Ev i
com
posi
Interaction Matrix
tion
cts
nfli
e co
Collective Solution with
conflicts
10
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Modeling context-aware distributed knowledge
Proposition: an incremental, distributed and incremental reasoning process
The composition of the interaction
matrix is done through the NPNe
Methodology
of
Aggregating
Cognitive Maps [Louçã,02a].
Only the most acute opinions are
considered to compose the
collective solution. A link such as
(+,--) evidences a conflict in the
organization. This way, the
collective reasoning mechanism
will detect and evidence conflicts
in
the
collective
solution,
graphically represented in the form
of a cognitive map, allowing a
clear discussion and negotiation
between the actors.
Individual solution
to a goal
Goal allocation : results from other reasoning processes
Solution
to an allocated goal
Solution
to an allocated goal
Solution
to an allocated goal
----------------Cognitive Map Aggregation-------------------------
Collective Solution
Mat
rix
ce c
d en
Ev i
com
posi
Interaction Matrix
tion
s
lict
onf
Collective Solution with
conflicts
11
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Modeling context-aware distributed knowledge
Propositions: cognitive maps standing for mental states
Cognitive maps composed on one hand by concepts and by causal links between
those concepts, in a strictu senso way [Weik,79] and on the other hand to consider
the context of concepts, allowing some kind of inference
This way, we can define mental states from cognitive maps by getting their
content from the concepts being causally related to their context. More precisely,
a mental state is represented by a concept and its context.
In cognitive mapping terms we have :
(1) concepts can’t be understood without context,
(2) the context of a concept is composed by concepts that influence
and that are influenced by the main concept and by links between
those concepts,
(3) each concept is coupled to its context, which can be called a
scheme [Bougon & Komocar,94]
12
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Modeling context-aware distributed knowledge
Propositions: cognitive maps standing for mental states
t6 - Create
propositions
+
e1 - Age of
employees :
30 / 35
+
(-,++)
e2 - Employees
notion of
"group"
t1 - Interact
(- -,++)
(+,++)
t2 - Internal
debates
b2 - Inovate
working
processes
++
++
t8 - Invest in R & D
+
t9 - Professionnal
learning
+
++
t7 - Research to
improve working
processes
e3 - Employees
confidence
b1 - Adapt
employees to
changes
(-,+)
--
t3 - Accept
suggestions
+
t4 - Periodical
reunions
+
+
t5 - Dialogue
e4 - Resistance to
change working
habitudes
Scheme – a concept and its context [Louçã,02]
13
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Modeling context-aware distributed knowledge
Propositions: cognitive maps standing for mental states
Structure Cognitive
Modèle Social
Modèle du Domaine
Module d'Analyse du
Domaine
Structure de
Raisonnement et
de Communication
Module de Résolution
to interact agents communicate schemes
communicating schemes influence agent’s mental states and
develops collective knowledge
Module d'Analyse
Sociale
Scheme
Module de
Communication
e1 - Age du
e2
personnel :
30 / 35
Transfert de données
e3
- Cohésion
du personnel
-
- Confiance
du personnel
Flux de Contrôle
+
t1 - Interagir
+
++
t2
- Faire
réflexion
interne
t9
- Faire formati
professionnelle
+
b1 - Adapter le
++
personnel au
changement
Structure Cognitive
---
- Accepter
suggestions
- Faire des
réunions
périodiques
+
+
t4
+
Modèle Social
Modèle du Domaine
t3
t5
e4
- Dialoguer
- Résistances à
changer d'habitudes
de travail
Module d'Analyse du
Domaine
Structure de
Raisonnement et
de Communication
Module de Résolution
Transfert de données
If concepts are similar and if there are
conflicts, then the graphical interface supports
discussion (and argumentation) between
individuals, to converge to a common scheme
Module d'Analyse
Sociale
Module de
Communication
Flux de Contrôle
Différents niveaux de détail dans la
visualisation d’une solution collective
avec mise en évidence de conflits
If concepts are not
similar, then
“CM aggregation”
14
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Application
Prototype and application in an industrial entreprise
These propositions above were tested in StrAgent, a distributed software system
to support decision-making in human organizations
SETCOM – Electronica, S.A. (Palmela)
industrial enterprise in the domain of telecommunications and electronics
Cognitive Maps were built from documents and interviews
5 teams - organisation, products, markets, human resources and information
departments
Main goal : to model collective discussion and consensual decision
• to represent actors knowledge
• to show conflicts
• to support conflict resolution
15
AAAI Spring Symposium 2003 on
AgentAgent-Mediated Knowledge
Management (AMKM 2003)
San Francisco,
Francisco, 24
24-26 March 2003
2003
Conclusion
A model of multi-dimensional reasoning in a multi-agent system is
proposed.
Cognitive maps are used as instruments to represent agent’s mental
states.
Schemes are used to compose a collective solution to a goal through a
distributed and incremental process, based on agent’s interactions.
The emergence of collective knowledge is represented in the cognitive
map of each agent.
16