Usually, ice sheets are reconstructed using the flow law of ice and a number of assumptions regarding ice sheet temperature, accumulation rate, and basal sliding resistance. In this paper, an alternative and independent reconstruction method is given. The weight of the ice sheets deformed the earth's surface and its geoid. The amount of deformation, which can be determined from presently emerged shorelines, is related to the thickness of the overlying ice burden. Therefore elevation and age data of sea levels from North America during the past 17,000 yr can be inverted to determine the history of the most recent North American ice sheet (Laurentide ice sheet), assuming the earth behaves as a viscoelastic material. Because of errors in the data and a number of simplifying assumptions, quadratic programing methods were necessary to insure that the predicted ice thicknesses of the past were always greater than at present. The predicted ice sheet is relatively thin (2000–3000 m), which suggests that it was very dynamic and probably not in a steady state. But poor resolution indicates that other ice thickness models, perhaps of greater thickness, may fit the sea level data as well. The fit to the data is, in general, good, and it indicates that a large proportion of the North American sea level data can be explained by glacial isostatic processes. However, at regions distant from the ice sheet (e.g., Florida) the fit is poor, because the assumption that water‐loading effects can be ignored is no longer valid. For regions close to a fluctuating ice sheet margin (e.g., Connecticut) the fit to the data suffers because of the assumption that the areal extent of the ice sheet remained fixed through time. Methods are developed which obviate both of these unrealistic assumptions so that future inversion calculations will be more realistic. Furthermore, proglacial lake deleveling data can be used to extend the data set throughout the interior of North America so that the entire perimeter of the ice sheet is close to the constraining data. The error of fit and resolution may thus be improved considerably.