Ever since the M(w)7.4 Izmit earthquake in 1999, evaluation of seismic hazard associated with the last unbroken segments of the North Anatolian fault is capital. A strong controversy remains over whether Marmara fault segments are locked or are releasing strain aseismically. Using a Bayesian approach, we propose a preliminary probabilistic interseismic model constrained by published GPS data sets. The posterior mean model show that Ganos and Cinarcik segments are locked while creep is detected in the central portion of Marmara fault. Our analysis, however, reveals that creeping segments are associated with large model uncertainties, which mainly results from the sparsity of current geodetic observations. We then discuss how the GPS network can be improved to attain more reliable assessment of interseismic slip rates. With this purpose, we implement a network optimization procedure to identify the most favorable distribution of stations measuring strain accumulation in the Marmara Sea.