Distributed game-theoretic topology control in cognitive networks

Existing distributed approaches to topology control are poor at exploiting the large configuration space of cognitive radios and use extensive inter-node synchronization to aim at optimality. We have created a framework to design and study distributed topology control algorithms that combine network-formation games with machine learning. In our approach, carefully designed incentive mechanisms drive distributed autonomous agents towards a pre-determined system-wide optimum. The algorithms rely on game players to pursue selfish actions through low-complexity greedy algorithms with low or no signaling overhead. Convergence and stability are ensured through proper mechanism design that eliminates infinite adaptation process. The framework also includes game-theoretic extensions to influence behavior such as fragment merging and preferring links to weakly connected neighbors. Learning allows adaptations that prevent node starvation, reduce link flapping, and minimize routing disruptions by incorporating network layer feedback in cost/utility tradeoffs. The algorithms are implemented in Telcordia Wireless IP Scalable Network Emulator. Using greedy utility maximization as a benchmark, we show improvements of 13-40% for metrics such as the numbers of disconnected fragments and weakly connected nodes, topology stability, and disruption to user flows. The proposed framework is particularly suitable to cognitive radio networks because it can be extended to handle heterogeneous users with different utility functions and conflicting objectives. Desired outcome is then achieved by application of standard cooperation techniques such as utility transfer (payments). Additional cross-layer optimizations are possible by playing games at multiple layers in a highly scalable manner.