Wirelessly-connected users encounter a vast array of environmental settings from factors including diverse terrain, buildings, vegetation, weather conditions, and velocities. Today, when traversing across these diverse conditions, each environmental change triggers a new wireless channel characterization so that links can have optimal performance for transmission rate and frequency band decisions, both of which depend on spatial and environmental characteristics. This characterization process can induce a high overhead on the network, greatly reducing the overall performance of the wireless links which were seeking to be optimized. In this project, diverse wireless scenarios will be classified into a finite set to recognize wireless channel types and optimize per-link and network-wide decisions. The project will significantly reduce the amount of characterization that needs to be performed per environment, especially when revisiting a location or when a new location shares many similarities as those previously visited.
The researchers will use two different approaches to classify wireless channels and create a notion of channel type which will be fed into an online training framework to optimize transmission rate and frequency band selection. In the first approach, previously encountered channel types will be recognized for immediate link-level and network-wide decisions and the resulting performance from these decisions will be observed to improve future decisions. In the second approach, previously-unencountered channel types for which insufficient levels of training exist will be inferred via crowdsourcing as an initial starting point for on-the-fly training. There are four main intellectual thrusts to the proposal: (i) The geometric relationship that forms between the link-level performance and the n-dimensional space of environmental factors contributing to such performance will be characterized in order to classify and recognize channel types across geographically diverse regions. (ii) A crowdsourcing approach with a large data set of cellular and WiFi-based mobile phone users will be leveraged to experimentally isolate the roles of geolocation, land use, and situational context to geo-spatially infer channel types. (iii) An on-the-fly training framework will be developed to leverage these two notions of channel type for previously-encountered and well-trained scenarios for optimal link adaptation, network-wide decisions, and ongoing training, and previously-unencountered or poorly-trained scenarios to use inferred channel types with decisions and resulting performance used for constructing a sufficient training. (iv) This knowledge of channel type will be exploited along with the current spectral activity and demand to optimize the band assignment in multiband, multihop networks.
