Realistic human behaviour simulation for quantitative ambient intelligence studies

Smart Homes technologies development is oriented toward intelligent services for the dweller. Designing the Artificial Intelligence which plays behind the scenes in a Smart Home requires large datasets for several reasons: training machine learning algorithms, tuning parameters, system testing and validation. Usually such tasks are carried-out on real-world data, requiring long time and additional costs to be collected, checked and labeled. Accelerating the development and limiting costs, a behaviour simulator can digitally reproduce environments and behaviours of the dwellers, in controlled conditions and in short time. This work presents a simulator capable of generating or reproducing the routine of a person in terms of Activities of Daily Living (ADLs). Moreover, the activity scheduling can be used to generate synthetic data from sensors deployed in a virtual environment. For the ADL schedule generation, an innovative model based on the person status (represented by needs) and habits is used, while two alternatives are proposed to generate home automation data: an agent-based model (with deterministic behavioural pattern descriptions) and a stochastic one (modeling the ambient response based on sample data activations distributions). The whole simulation/emulation chain is evaluated comparing the characteristics of the obtained data with a real world dataset. This comparison proves that synthetic data respect the distributions of the corresponding real world dataset ADLs and sensors activations.