Regional source‐based earthquake early warning systems perform three consecutive tasks: (1) detection and epicenter location, (2) magnitude determination, and (3) ground‐motion prediction. The correctness of the magnitude determination is contingent on that of the epicenter location, and the credibility of the ground‐motion prediction depends on those of the epicenter location and the magnitude determination. Thus, robust epicenter location scheme is key for regional earthquake early warning systems. Available source‐based systems yield acceptably accurate locations when the earthquakes occur inside the real‐time seismic network, but they return erroneous results otherwise. In this study, a real‐time algorithm that is intended as a supplement to an existing regional earthquake early warning systems is introduced with the sole objective of ameliorating its off‐network location capacity. The new algorithm combines measurements from three or more network stations that are analyzed jointly using an array methodology to give the P‐wave slowness vector and S‐phase arrival time. Prior to the S‐phase picking, the nonarrival of the S phase is used for determining a minimum epicentral distance. This estimate is updated repeatedly with elapsed time until the S phase is picked. Thus, the system timeliness is not compromised by waiting for the S‐phase arrival. After the S wave is picked, an epicentral location can be determined using a single array by intersecting the back‐azimuth beam with the S‐minus‐P annulus. When several arrays are assembled, the back azimuth and P and S picks from all arrays are combined to constrain the epicenter. The performance of the array processing for back azimuth and S‐wave picking is assessed using a large number of accelerograms, recorded by nine strong motion sensors of the KiK‐net seismic network in Japan. The nine stations are treated as three distinct seismic arrays, comprising three stations each. Good agreement is found between array‐based and catalog‐reported parameters. Finally, the advantage of the new array methodology with respect to alternative schemes for back azimuth and distance is demonstrated.