Learning with play behaviour in artificial agents

Authors Organisations
Type

Student thesis: Doctoral ThesisDoctor of Philosophy

Original languageEnglish
Awarding Institution
Supervisors/Advisors
Award date2019
Links
View graph of relations

Abstract

This thesis investigates the role of exploratory play in the development of the basic knowledge of actions and objects involved in play with artificial agents. We developed a learning system, Dev-PSchema, inspired from the sensorimotor stage and schema mechanism of Piaget’s theory of cognitive development. The learning system enables the artificial agents to develop their knowledge, in the shape of schemas containing action and perceptions, based on their existing knowledge and play behaviour. We demonstrate the system embodied in
two agents, a simulator and a real robot, developing their knowledge through exploratory sensorimotor experiences and extending for novel situations of the environment through schema generalisation. The schema generalisation mechanism enables the agents to extend their knowledge for novel situations and predict action outcomes. The agents begin learning with a set of basic actions, provided to interact with their environment and perform suitable actions selected through an action selection mechanism, the excitation calculator.
This mechanism is modelled on the habituation paradigm, widely studied in developmental psychology. We demonstrate how the excitation mechanism can be tuned to demonstrate a range of behaviour preferences in the artificial agents, similar to the infants observed in the developmental psychology studies.
We then demonstrate the agents developing complex actions, labelled as schema chains, from their basic knowledge gained through the exploratory play. The agent demonstrates performing the schema chains as a singular action following a few repetitions. This skill is modelled on the chain reflex and motor program behaviours observed in humans while performing a sequence of actions. This capability enables the agent to develop complex skills that are used to achieve a state in the environment which is, otherwise, not possible to achieve with a single action. Furthermore, we demonstrate the capability to scaffold the
learning of the agent through achieving tasks, increasing in complexity.
In conclusion, we demonstrate that the exploratory play helps the agents to develop their knowledge about actions and the objects involved. The developed knowledge is further used to explore the environment, hence demonstrating open-ended learning

Keywords

  • developmental robotics, modelling play behaviour, artificial intelligence