Developing the Reinforcement-Learning Child Agents for Measuring Play and Learning Performance in Kindergarten Design

초록

Although kindergarten design requires promoting play behaviours related to physical and social development and ensuring safety of children to support a child-oriented design, the systemic evaluation of design performance remains challenging because of the children’s dynamic play behaviours. As one solution, agent-based simulations have been applied in the design field, but there are limitations in reflecting children’s behavioural richness. To overcome this shortcoming, this study developed reinforcement learning (RL) child agents to compute the dynamic play behaviours associated with physical and social development. Several iterations were conducted to implement the RL agents’ play behaviours, and the results were incorporated into the simulation. To validate the play behaviour model, we conducted a case analysis with authentic and unbuilt kindergarten designs and measured the quantifiable design performance in terms of physical and social play behaviours and safety. The results indicated that the RL child agents enabled a holistic analysis and the calculation of generative behavioural responses, depending on physical variations. By facilitating an unknown design affordance, the RL-powered simulation model is expected to provide data-driven evidence to support a child-oriented design.