The problem of estimating cosmological parameters such as from noisy or incomplete data is an example of inverse problems and, as such, generally requires a probabilistic approach. We adopt the Bayesian interpretation of probability for such problems and stress the connection between probability and information which this approach makes explicit. This connection is important even when information is `minimal' or, in other words, when we need to argue from a state of maximum ignorance. We use the transformation group method of Jaynes to assign minimally-informative prior probability measure for cosmological parameters in the simple example of a dust Friedmann model, showing that the usual statements of the cosmological flatness problem are based on an inappropriate choice of prior. We further demonstrate that, in the framework of a classical cosmological model, there is no flatness problem.