Thursday, April 23, 15:30 - 18:00, Location: Show and Tell Area B
Robert C. Daniels, Ketan Mandke, Steven W. Peters, Scott M. Nettles, and Robert W. Heath, Jr.
This demonstration presents a novel approach to physical layer link adaptation, or data rate selection, through online machine learning. The implementation uses Hydra, the IEEE 802.11n draft standard multihop wireless networking prototype. The demonstration highlights both the utility of learning-based link adaptation in practical networks as well as the flexibility of Hydra.
Link adaptation is a technique for adapting the physical layer transmission of packets to take advantage of variations in the wireless channel. For example, this is done in Hydra by modifying the modulation and coding scheme. In practice, link adaptation is especially challenging due to intractable complexities in the physical layer performance due to the combined use of multiple antennas, convolutional codes, and orthogonal frequency division multiplexing (OFDM). Recently, it has been shown that machine learning provides a powerful infrastructure for link adaptation. Machine learning techniques are very attractive since they allow communication systems to discover the effective link adaptation techniques online. This is a paradigm shift from the current link adaptation practice where link adaptation procedures are calculated offline. Online learning is attractive for several reasons in wireless communication systems including:
* Offline measurements used to determine link adaptation strategies suffer from degraded performance when changes occur during the lifetime of a system. These changes could be, for example, modified front end filtering and sampling characteristics or performance degradation with heated analog circuitry.
* Consumer devices are manufactured with high yield to meet the demands of the current wireless market. Although each device meets certain performance requirements, each system performs uniquely due to subtle differences in analog circuits. However, link adaptation code generated offline cannot be tuned to each unique wireless device. Hence, online learning with link adaptation can be used in each system to maximize the unique performance of each system.
The adaptation procedure is described below in six-steps:
1) On each link of the multihop prototype, the MAC layer of the transmitting node (TX Node) will probe the channel to the intended destination.
2) The receiving node (RX Node) receives the probing message from the TX Node and calculates channel quality measurements which are transformed into feature spaces parameters. The current feature space realization queries the feature space database.
3) A classification algorithm determines the highest possible rate that will be successful given the current feature space realization. This rate (or modulation and coding scheme) is sent to the RX Node.
4) Over-the-air feedback sends this rate value from the RX Node to the TX Node.
5) The TX Node opportunistically adapts the physical layer parameters for a transmitted data packet to optimize rate.
6) The RX Node determines if the data packet was correctly received. The feature space realization and whether or not it was correctly received at the RX Node is entered into the feature space database to facilitate online learning.
Our proposed of Hydra will also utilize various GUIs to visualize variations in channel state, PHY layer adaptation, and trace MAC and application level messages. These visualizations provide great intuition about the distribution of the real propagation channel, the bottlenecks in system design, and the benefits of the cross-layer protocols implemented. The advantage of experimentation with prototypes such as Hydra is that details about channel distributions, overhead, traffic distributions, hardware impairments, etc. cannot be neglected. Such experience can give engineers keen insight into system design, help to distill the importance of various research questions, and even drive the direction of research.