This paper proposes a simulator for reproducing indoor human movements to facilitate developments and validations of ambient assisted living (AAL) systems. The proposed simulator learns actual behavior logs of real people to generate a graph data structure that represents their life interests. These generated graphs are then assigned to virtual agents in the simulator to make them behave similarly as real people in the simulated environments. Compared to conventional ones, the three main contributions of our proposed simulator are: (1) It is scalable to simulate indoor movements of a large number of people from a limited number of actual behavior logs; (2) It is adaptive to simulate how differently people would behave in different arrangements and layouts of the environment; and (3) It is flexible to simulate how people would behave when they face an unexpected situation. This paper mainly focuses on the conceptual justification of the proposal and we have validated novelties of the proposed simulator from its design and implementation. Finally, perspectives on its quantitative evaluation have also been discussed thoroughly.