In this letter, we propose a novel approach for fusing information from multiple access points in order to enhance WiFi-based self-localization. A common approach for designing WiFi-based localization systems is to learn location-to-signal strength mappings for each access point in an environment. Each mapping is then used to compute the likelihood of the robot's location conditioned on sensed signal strength data, yielding as many likelihood functions as mappings are available. Office buildings typically have from several tens to a few hundreds of access points, making it essential to properly combine all available likelihoods into a single, coherent, joint likelihood that yields precise likelihoods, yet is not overconfident. While most research has focused on techniques for learning these mappings and improving data acquisition; research on techniques to adequately fuse them has been neglected. Our approach for data information fusion is based on information theory and yields considerably better joint distributions than previous approaches. Furthermore, through extensive testing, we show that these joint likelihoods considerably increase the system's localization performance.