Use randomness and select a subset of all hypotheses. i.e. Monte Carlo method. Perhaps use an evolutionary algorithm that that starts with seed hypotheses and vary the ones that get closest to the observation