Many devices are nowadays equipped with multiple wireless network interfaces such as GSM, HSPA+ and WLAN. This requires a mechanism for identifying which access network to use at a given time so to provide best user experience. Such decision is typically made using past information available at the user (e.g. history of signal strength measurements), or using a prediction of the future status of the parameters of interest. The key problem is that the parameter estimation is difficult and that the predictions may be inaccurate. In this paper, we develop a mathematical model for the access network selection problem, which is solved using robust optimization techniques. The objective is to maximize the download rate the user can achieve during a certain period in the future. The model provides guidelines for selecting the access network which guarantees, within a given probability bound, a minimum download rate for a given rate uncertainty.