The usual approach to ambient intelligence is an expert modeling of the devices present in the environment, describing what each does and what effect it will have. When seen as a dynamic and noisy complex systems, with the efficiency of devices changing and new devices appearing, this seems unrealistic. We propose a generic multi-agent (MAS) learning approach that can be deployed in any ambient environment and collectively self-models it. We illustrate the concept on the estimation of power consumption. The agents representing the devices adjust their estimations iteratively and in real time so as to result in a continuous collective problem solving. This approach will be extended to estimate the impact of each device on each comfort (noise, light, smell, heat...), making it possible for them to adjust their behaviour to satisfy the users in an integrative and systemic vision of an intelligent house we call QuaLAS: eco-friendly Quality of Life in Ambient Sociotechnical systems.