Adaptive learning approaches for smart home environments with a simulator implementation
Smart home solutions are utilized in a different way by each user according to the user’s unique needs and preferences over his unique home setting. User – device interactions themselves carry some implicit characteristics of the context the smart devices are used. In order to better exploit Internet of Things technologies in smart home environments, the user’s interaction history can be leveraged to generate knowledge of his behavior habits. With a purpose of achieving personalized decision making of smart lighting environments, this thesis presents three learning model approaches, which are based solely on the individual’s interaction habits with the devices, adaptive to changes in inhabitant’s device usage behavior, and do not require preset data for initialization. Individual interactions with smart light devices are contextualized based on the timestamp of the interaction, and ambient light intensity reading of the room at that instant. Moreover, the thesis introduces a Java simulator, which interactively demonstrates the behavior of a personalized smart home setting with the integrated learning models with a feedback mechanism. The simulator leads up to a way of evaluating adaptive learning models in smart home environments, reducing the immediate need of testing their behavior in a real life smart home, which is both hard and costly to construct and maintain. The learning approaches, namely K-Nearest Neighbor and Softmax Regression with batch learning and online learning algorithms, are evaluated with different datasets representing different scenarios, and the results show that all three methods are able to perfectly capture the usage pattern when the interactions with distinct “things”, smart light devices, are separable in terms of their corresponding context. For more complex datasets, which have overlaps between the usage context of distinct devices and big changes in user behavior over time, the online learning algorithm needs more data in order to catch the performance of KNN and Softmax Regression with batch learning algorithms.