Designing Agents to Understand Infants

Dean Petters

2006 · 17 citations · 87 references

Concepts

Abstract

This thesis attempts to understand infant behaviour by designing autonomous software agents to reproduce those behaviours in simulation. The infant behaviours that have been investigated are related to the phenomenon of infant attachment. Empirical studies of infant behaviour are abstracted and the function of the behaviours are assessed from an evolutionary perspective. The behaviours are then reformed as scenarios against which simulations can be evaluated. These studies include naturalistic observation of infants at home and exploring a park, and undergoing the Strange Situation Experiment. A number of information processing architectures have been constructed that reproduce the infant behaviours described in scenarios. These vary in complexity from a reactive architecture with no capacity to learn, to reactive architectures that can learn by reinforcement, and deliberative architectures that can reason by forming simple plans. Computational experiments undertaken with interacting infant and carer agents show the presence of interesting dynamic properties, such as positive feedback loops. These feedback loops may provide an explanation for the empirical finding that patterns of infant attachment response cluster into three categories. This thesis demonstrates how the methodology that it is uses in investigating attachment behaviour in infancy may be extended to many other infant and developmental behaviours.

References

87