Machine Learning for MIMO Link Adaptation
MIMO-OFDM (multiple input multiple output orthogonal frequency division multiplexing) is becoming the de facto transmission technique for many commercial wireless systems, e.g. IEEE 802.11n, IEEE 802.16, and 3GPP. Despite the broad acceptance of MIMO-OFDM, leveraging the capacity and reliability offered by MIMO has been challenging due to the extra adaptation lever introduced by the spatial dimension. This lever can be used, for example, to switch between multiplexing (high rate) and diversity (high reliability) modes of operation. Adapting to maximize throughput is challenging because the optimum transmission mode depends on several features of the channel, not just the received SNR as in conventional non-MIMO systems. Even with complete channel state information, mathematical expressions for choosing the best parameters may not be available, e.g. with coding and interleaving. If such expressions are available, they usually rely on idealistic assumptions about the system and neglect impairments like estimation error, phase noise, synchronization mismatch, or interference. Consequently, most prior practical work on link adaptation used some combination of lookup tables or other ad hoc approaches like autorate fallback.
Recently, we developed a new paradigm for link adaptation based on concepts from machine learning. The idea is that the receiver can learn how to optimize performance by observing and remembering the consequences of its previous decisions. In our initial work, we developed an algorithm for link adaptation in convolutionally coded MIMO-OFDM systems. Using an offline training phase, we were able to create a channel mapping to maximize throughput subject to a target error rate constraint. A major insight in that work was that certain quantiles of the ordered post-processing SNR could be used to build an accurate feature vector. More recently, we developed an online database for link adaptation in MIMO-OFDM systems. The receiver maintains a database of previous decisions and their outcomes. Then, it makes a channel measurement and queries the database to recommend a rate for the current channel. An advantage of the online approach is that it can adapt to different channel conditions in real-time and makes minimal assumptions about the operating environment. We have implemented our work on Hydra.
Several of our recent publications are summarized below.
R. Daniels and R. W. Heath, Jr., “Online Adaptive Modulation and Coding with Support Vector Machines,” (invited) Proc. of the IEEE European Wireless Conference , Lucca, Italy, pp. 718-724, April 12-15, 2010.
This paper presents an online learning approach for rate adaptation using support vector machines. Compared with our previous approach using the nearest neighbor algorithm, support vector machines provide a more practical path to efficient low-complexity reduced-memory implementation.
R. Daniels and R. W. Heath, Jr., “Link Adaptation in MIMO-OFDM with Non-Uniform Constellation Selection over Spatial Streams through Supervised Learning,” Proc. of the IEEE Int. Conf. on Acoustics, Speech, and Signal Proc., Dallas, TX, pp. 3314-3317, March 14-19, 2010.
This presents a lower complexity approach to link adaptation. It allows for potentially different constellation sizes for each MIMO data stream. We show that the proposed approach has lower complexity and somewhat higher performance than our previous approach that fixed the rate per-stream.
R. Daniels, C. Caramanis, and R. W. Heath, Jr., “Adaptation in Convolutionally-Coded MIMO-OFDM Wireless Systems through Supervised Learning and SNR Ordering,” IEEE Trans. on Veh. Tech., vol. 59, no. 1, 2010, pp. 114-126.
This paper proposes a new methodology for link adaptation based on a concept from machine learning known as supervised learning. The work in this paper was inspired by our observation that link adaptation in real MIMO-OFDM systems is hard. Supervised learning provides a way to implement link adaptation in a MIMO-OFDM including spatial mode adaptation, modulation selection, and code rate selection through performance-related observations.
R. Daniels and R. W. Heath, Jr., “An Online Learning Framework for Link Adaptation in Wireless Networks,” Proc. of the Information Theory and Applications Workshop, San Diego, CA, February 8-13, 2009, pp. 138-140.
This paper proposes an online approach to machine learning for rate adaptation. Compared with supervised learning, this online framework uses real-time measurements to update the rate-adaptation classifier. There are many interesting tradeoffs with online implementations including exploration vs. exploitation (how often should the system explore new rates to keep the mappings up to date), and database management (how should the database tradeoff new data versus diverse data). Online learning has been implemented and demonstrated on Hydra.
We won the Grand Prize in the 2008 WinCool Demo Contest, held in conjunction with the Third International Workshop on Wireless Network Testbeds, Experimental Evaluation, and Characterization, on conjunction with ACM MobiCom 2008. The prize was one for a demonstration of online learning for rate adaptation using Hydra.
We have been fortunate to have several outstanding sponsors support this work including the National Science Foundation under grant CNS-626797, and US Army Research Laboratory under grant number W911-NF-08-1-0438, and the DARPA IT-MANET program, Grant W911NF-07-1-0028. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the aforementioned sponsors.