Observed time delays between images of a lensed QSO lead to the determination of the Hubble constant by Refsdals method, provided the mass distribution in the lensing galaxy is reasonably well known. Since the two or four QSO images usually observed are woefully inadequate by themselves to provide a unique reconstruction of the galaxy mass, most previous reconstructions have been limited to simple parameterized models, which may lead to large systematic errors in the derived H_0 by failing to consider enough possibilities for the mass distribution of the lens. We use non-parametric modeling of galaxy lenses to better explore physically plausible but not overly constrained galaxy mass maps, all of which reproduce the lensing observables exactly, and derive the corresponding distribution of H_0s. Blind tests - where one of us simulated galaxy lenses, lensing observables, and a value for H_0, and the other applied our modeling technique to estimate H_0 indicate that our procedure is reliable. For four simulated lensed QSOs the distribution of inferred H_0 have an uncertainty of simeq 10% at 90% confidence. Application to published observations of the two best constrained time-delay lenses, PG1115+080 and B1608+656, yields H_0=61 +/- 11 km/s/Mpc at 68% confidence and 61 +/- 18 km/s/Mpc at 90% confidence.