Network Selection in Cognitive Radio Systems Chonggang Wang ∗ , Kazem Sohraby † , Rittwik Jana ‡ Lusheng Ji ‡ Mahmoud Daneshmand ‡ ∗ NEC Laboratories America, Princeton, New Jersey 08540, USA † University of Arkansas, Fayetteville, AR 72701, USA ‡ AT&T Labs Research, Florham Park, New Jersey 07932, USA Abstract— Measurement studies have shown that uneven and dynamic usage patterns by the primary users of license based wireless communication systems often lead to temporal and spatial spectrum underutilization. This provides an opportu- nity for secondary users to tap into underutilized frequency bands provided that they are capable of cognitively accessing without colliding or impacting the performance of the primary users. When there are multiple networks with spare spectrum, secondary users can opportunistically choose the best network to access, subject to certain constraints. In cognitive radio systems, this is referred to as the network selection problem. In this paper, multiple network selection strategies namely, random, weighted, and greedy, are comprehensively evaluated. It is found that without adequate admission control, those methods cannot provide sufficient service protection for the primary users. Next, a Markov decision model is applied to obtain the maximum allowable arrival rate for secondary users subject to a target collision probability for the primary users. Based on this model, a Collision-Constrained Network Selection (CCNS) method is proposed that maximizes system throughput subject to a given collision probability. Simulations show that comparing to random, weighted, and greedy strategies CCNS achieves an improved performance in terms of system throughput and collision probability. I. I NTRODUCTION Demand for wireless network services has increased dra- matically. This trend is expected to continue in the coming years which necessitates allocation of more radio spectrum to support these services. Most of the radio frequency spectrum under 3 GHz is already exclusively assigned. Several wireless services are now constrained to use overpopulated unlicensed bands (e.g. ISM) or bands beyond 3 GHz [1], which are considered less than ideal for wireless operators due to radio propagation characteristics at high frequencies. Historically, spectrum resources have been statically reg- ulated: different frequency bands are allocated for different usage and the right to use such frequency band is licensed. Independent measurement studies have shown significant un- derutilization in various licensed frequency bands. This obser- vation has prompted several regulatory, technical and legisla- tive bodies to rethink spectrum allocation for a more efficiently managed spectrum [1]. To solve this scarcity problem, technologists conceived the idea of improving spectrum usage by opportunistically allowing secondary users to utilize unused licensed bands, typically referred to as the “spectrum holes” [2], provided that additional interference is not introduced by secondary users [3]. With this approach, the secondary users must employ special devices known as Cognitive Radios (CR) which are capable of sensing their RF environment and modifying its communication configurations for accessing “spectrum holes”. Such an approach has interested both the Federal Communi- cation Commission (FCC) [4] and standards bodies such as IEEE and ITU-R [5]- [8]. Sharing of licensed spectrum in CR system is “verti- cal” [9]due to differences in priorities in spectrum usage between primary users (PUs) of the network and opportunistic secondary users (SUs). If all channels 1 are occupied by PUs and SUs, a newly arriving PU will cause an existing SU to be evicted from the network, which is referred to as “collision”. Admission control and scheduler designs such as those for the packet-level schemes proposed in [10] target the overall sys- tem throughput maximization subject to certain collision rate. When a CR system incorporates multiple primary networks, SUs also face the problem of choosing which primary network to join. This is known as the Network Selection Problem (NSP) in CR. [11] studied the problem of network selection in the context of a multi-commodity optimization problem. This current paper studies the network selection problem from the perspective of “collision probability”. In fairness, a complete spectrum management framework needs to be instituted so that it deals with all of the above challenges [12], such as: 1) interference avoidance, 2) spectrum sensing, 3) spectrum decision, 4) spectrum sharing, and 5) spectrum mobility. In general, the optimization of cognitive network selection strategy requires the presence of a centralized controller that has complete knowledge of all primary networks and interfer- ence structure at every receiver. [13] studied the problem of a decentralized selection strategy where a Nash equilibrium is reached amongst SUs. However, the algorithms proposed in this paper apply to both centralized and decentralized selection strategies. The major contributions of this paper are as follows: • Comprehensive evaluation of three network selection strategies: random, weighted, and greedy. It will be shown that without admission control, none of these algorithms guarantee a desired limit on the collision probability, al- though weighted and greedy perform better than random in terms of collision probability performance. • Establishing a Markov model and quantitatively calcu- lating collision and PU/SU blocking probabilities in a multi-network environment. • Based on the above Markov model, we design a 1 Here we use the term “channel” in a logical sense. A channel is a dedicated allocation of resources for a user’s communication needs. This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE "GLOBECOM" 2009 proceedings. 978-1-4244-4148-8/09/$25.00 ©2009