Dean Petters

1

Dean Petters
Using Software Agents to Simulate Infant
Secure-base Behaviour
Simulation 1 shows patterns of naturalistic
infant exploration, observed with infants moving
around stationary carers in London parks
(Anderson, J.W. Attachment behaviour out of
doors, in Blurton-Jones, 1972). The traces left
by the simple simulated infants are straight and
symm-etrical because the software infant agents
do not currently possess mechanisms that allow
them to be distracted.
This paper presents a number of autonomous agent
software simulations as theories of infant
attachment. The process of creating each
simulation involves several stages selecting
empirical observations abstracting them into the
form of a scenario which acts as a specification
of requirements forming a design and
implementing and then evaluating that design in
relation to its scenario.
Design 1 A hybrid infant cognitive architecture
(with reactive and del-iberative components)
designed and implemented and which can reproduce
the behaviours seen in simulations 1 to 5. In
the deliberative subsystem information transfer
is discrete and on-demand, in comparison with the
cont-inually updated transfer in reactive lower
processes.
Simulations 3 to 5 reproduce infant behaviour in
the Strange Situation
(b) Toys out-side safe-range so infant moves out
to explore, gets anxious, and returns to carer
Simulation 5 shows an Insecure-Ambivalent infant
with a small safe-range. The carer is outside of
the Safe-range and the infant moves and signals
to the carer (signalling represented by going
purple)
Simulation 3 shows an Insecure-Avoidant infant
with a small safe-range resulting from
unresponsive caregiving. The carer is outside of
the infants safe-range and the infant is anxious
and moves to the carer but does not signal
Simulation 4 shows a Secure infant. This infant
has a large safe-range that resulted from a
history of responsive caregiving.The carer is
inside the Safe-range and the infant moves
towards a toy to play
(a) Toys within safe-range so infant explores all
toys without turning to carer
infant
infants Safe- range
infants route
carer
infant
Small safe range
toy (object to explore)
infant
infant
(d) Anxiety rises faster than (b) so infant never
gets to toys before returning
Large safe range
Small safe range
infantvisual range
carer
carer
Infants most active internal state anxiety
Infants most active internal state anxiety
Infants most active internal state explore goal
carer
infantvisual range
infantvisual range
(c) Anxiety drops slower than (b) so infant
returns closer to carer before anxiety drops
lower than exploration
anx anxiety, obj object wariness, soc
social wariness
Experiments 1 to 4 Computational experiments
were carried out to investigate the relative
effects of two learning mechanisms. Firstly,
infants learned about their attachment
relationships solely from carer responses in
episodes when they experience anxiety
(goal-learning-from-anxiety, GLA architecture).
Secondly, infants learned solely from episodes
where they experience warmth or rejection in a
non-anxious environment (goal-learning-from-sociab
ility, GLS architecture). Results show that
learning from episodes where anxiety was
experienced causes greater clustering of infant
security responses (measured by the safe-ranges
that infants learn to keep their carers within).
This phenomena may explain part of the
empirically observed clustering seen when
measuring Secure and Insecure categories of
infant in Strange Situation Studies.
Simulation 2 shows a carer and stranger moving
from left to right as a coy infant repeatedly
approaches the stranger and then retreats to the
secure-base provided by the carer.
stranger
Start of coy episode
carer
Infant safe range
stranger
End of coy episode
Cycles in the simulation
Ten runs of the GLA simulation, run over ten
thousand cycles. For each run, the carers
responsiveness is set to give an intermediate
level of security. However, a positive feedback
loop accentuates initial random differences,
giving rise to two discrete clusters
Ten runs of the GLS simulation, run over ten
thousand cycles. For each run the carers
responsiveness is set to give an intermediate
level of security. Since there is no positive
feedback loop, no clusters are formed and a range
of levels of infant-carer security result
carer
Ten experiments with the GLA simulation, each
data point shows twenty individuals run over one
thousand cycles. There is evidence of a critical
point that splits secure from insecure
infant-carer pairs with few intermediate cases.
Ten experiments with the GLS simulation, each
data point show twenty individuals runs over one
thousand cycles. There is not evidence of a
single critical point that bifurcates secure and
insecure infant-carer pairs.
Note a limitation the stranger doesnt become
more familiar to the infant during the episode