Simulating Attachment
 Why simulate attachment?
 Origins of Attachment theory
 The target behaviours to be simulated
 Design methodology and demo
 Architectural design issues to be
investigated
Why Simulate Attachment?
It provides a target for a design
process - by building cognitive
architectures to perform certain
specific tasks we better understand
architectures generally
Reproducing in simulation specific
patterns of attachment behaviour acts
as a ‘test-bed’ for developing
architectural theories of human
information processing
Why Simulate Attachment?
the developmental trajectory has
normative stages which show
representational change
initially only need to simulate limited
cognitive resources
linked with evolutionary, physiological,
anthropological, AI, cybernetic and
cross-species data and theory
can abstract attachment behaviour
Origins of Attachment Theory
John Bowlby, The Attachment Trilogy
Psychoanalysis, Ethology, Evolutionary
Theory and Cybernetics
Early concentration on long
separations, loss, mistreatment and
psychopathology
Changed hospital visiting practice
Later focus on Individual Differences
The target behaviour
 The Strange Situation Experiment
arose from comparing Ugandan and US
infant attachment behaviours
 Involves 3 separation/re-union stages
 Each new stage increases the amount
of anxiety they produce
 Four patterns of response
 A key pattern is link between home
behaviour of mother and infant and
infant behaviour on re-union in the SS
The target behaviour
Infant reunion responses in the SS:
Secure (B type) behaviour
positive, greeting, being comforted
Avoidant (A type) behaviour
not seeking contact, avoiding gaze
Ambivalent (C type) behaviour
not comforted, overly passive, show anger
Disorganised (D type) Behaviour
totally disorganised and confused
The target behaviour
Maternal home behaviour
prior to the SS
sensitivity-insensitivity
acceptance-rejection
co-operation-interference
accessibility-ignoring
emotional expressiveness
rigidity(compulsiveness)-flexibility
The target behaviour
 Attachment SS Subgroups vs prior maternal
home behaviour
SS subgroups
Maternal Behaviour\
B1
B2/B3 C1 C2
Sensitivity
7.36 4.50
2.50 2.25 2.50 2.75
Acceptance
8.00 6.75
5.50 5.25 4.25 3.50
Co-operation 7.66 6.50
4.00 4.50 5.50 2.63
Accessibility 7.39 4.88
4.50 2.50 2.25 4.63
Lack of emotion* 2.17 3.88
4.25 2.75 6.50 6.00
Rigidity*
2.00 3.50 3.50 4.50
1.89 2.75
A2 A1
Design methodology
3 Problems:
 Avoiding trivial solutions
 Whether to use data or theory to
constrain the simulation
 Non-falsifiability
Design methodology
Problem 1:
Avoiding trivial solutions
 The simulation is NOT trying reproduce
superficial details of facial expression or
body movement - like a Kismet robot might
 It is trying to simulate the causal
mechanisms behind the behaviour, at the
level of goals and action plans within a
complete agent in a multi-agent simulation
 BUT an abstraction of the target
behaviour in a broad and shallow complete
agent is TOO easy to reproduce
Design methodology
Problem 1:
Avoiding trivial solutions
 How to constrain the
possible hypotheses space to
exclude trivial solutions?
 Assume attachment styles
are evolved, adaptive
behaviours
Design methodology
Problem 1:
Avoiding trivial solutions
Concentrating on Secure (B type),
Avoidant (A type) and Ambivalent (C
type) behaviours as potentially
adaptive responses
Disorganised (D type) Behaviour is
unlikely to be adaptive, but inclusion
of this phenomena remains a possible
future constraint
Design Methodology
 Taking an evolutionary/adaptive approach the differences in
infant security in the Baltimore and Uganda studies
suggests the following questions:
Are Internal Working Models that are used in moments of
attachment anxiety in part formed in episodes centred on nonanxious socialisation and exploration?
What information might infants gain from frequent episodes of
exploration and social interaction that they use in infrequent
episodes of attachment anxiety?
“If my carer won’t socially interact on my terms at all then I am less
secure and I must use my own actions to gain security”
“If my carer sometimes socially interacts on my terms then I am less
secure and need to concentrate my efforts in eliciting a response”
Design methodology
Problem 2:
How to combine data, theory and
AI techniques in the simulation - (Mook (1983) In
defense of external invalidity)
 Data and theories to be
incorporated in the simulation
 Data from the SS and other attachment studies
 Bowlby’s theory
 Distributed control architectures
 Teleoreactive architectures
 H-cogaff architecture
 Theories of Executive Function - SAS
 Machine learning algorithms (RL and ILP)
Design methodology
Problem 3:
Non-falsifiability
 Duhem, Auxiliary Assumptions
 Popper, Falsifiability and Broad and Shallow
architectures
 Lakatos,
 Three kinds of falsification
 Core assumptions and ad hoc assumptions
 Progressive and Degenerative Problem Shifts
Design methodology
An unfinished
simulation
Architectural design issues
how goals are chosen and represented?
when goals are chosen how are
consequent behaviours chosen?
whether SS behaviour is deliberative or
reactive?
how skill acquisition, chunking, parsing,
perceptual affordances relate to
attachment?
when and how new subsystems come
on-line in attachment stage changes?
Architectural design issues
Bowlby’s theory
Behavioural systems from ethology:
attachment, exploration, fear and sociability
Stages defined by available control
mechanisms: reflex (0-3), fixed action patterns
(2-12), goal correction (9-36), goal corrected
partnership (24- ), (age in months)
Coordination and control mechanisms:
chaining, planning, ‘totes’, IWM’s and language
Architectural design issues
The H-cogaff
architecture
Architectural design issues
Shallice and
Burgess - SAS and
contention
scheduling
SAS
contention scheduling
The cogaff schema
Bowlby’s Behaviours represented in an
infant-cogaff architecture
Internal
Working
Model?
Architectural design issues
Development
of deliberative
affordances or
exploration
and
socialisation
driven by
deliberation?
Development
of
partnership
in planning
as linguistic
competence
develops
Deliberation
in re-union
episodes
Architectural design issues
 A distributed control system that adapts with
Re-inforcement Learning at each node, has a
non-central, non-symbolic representation,
given by the genes, and undergoes no
qualitative change in representation
 A teleoreactive system that adapts using
Inductive Logic Programming, has a simple
central symbolic representation given by the
genes that undergoes qualitative change in
representation