Studies on agent design have been focused on the internal structure of an agent that facilities decision-making subject to domain specific tasks. The domain and environment knowledge of an artificial agent is often hard coded by system engineers, which is both time-consuming and task dependent. In order to enable an agent to model its general environment with limited human involvement, in this paper, we first define a novel autonomous hypothesis generation problem. Consequently, we present two algorithms as its solutions. Experiments show that an agent using the proposed algorithm can correctly reconstruct its environment model to a certain extent.