This thesis pursues the idea of creating a home or building that automatically adapts to the requirements and desires of its occupants.
To ensure that occupants are comfortable during various activities, it is often necessary to coordinate multiple operational parameters from systems belonging to different building disciplines. Furthermore, individual structural properties or the building are to be included into the control strategy. If there are multiple options available to realise the desired conditions, the cheapest method shall be chosen, but occupant comfort shall not be compromised. This work focuses on two points of research regarding the implementation of self-learning system behaviour for home automation systems. The first focus is on the question how artificial neural networks can be used to create an intelligent heating control loop that better considers the complex parameters and inputs coming from humans, buildings, building services, and the natural environment. The artificial neural networks created for this purpose are of the Feed Forward Network, Jordan Network, and Simple Recurrent Network type. According to a presence schedule, outdoor and indoor temperatures and the current time, they modify the set value of a conventional heating controller. The resulting increase in efficiency over a conventional, purely time based strategy is evaluated using physics-based simulation. The second focus is on the structuring of complex control tasks and self-learning capabilities within the framework of a software application. To this end, a multi-agent system is implemented. In it, various types of agents reflect the various learning and control tasks. Besides the agent types themselves, their dynamic interplay is also modelled.