Intel Seminar

Design of Interference-Aware Wireless Communication Systems

Prof. Brian L. Evans
Dept. of Electrical and Computer Engineering
Wireless Networking and Communications Group
The University of Texas at Austin, Austin, Texas

bevans@ece.utexas.edu

Lead graduate students: Mr. Aditya Chopra, Mr. Kapil Gulati and Mr. Marcel Nassar

Current collaboration with Dr. Nageen Himayat, Mr. Kirk Skeba, and Ds. Srikathyayani Srikanteswara at Intel

Past collaboration with Ms. Chaitanya Sreerama, Dr. Eddie X. Lin, Dr. Alberto A. Ochoa, and Mr. Keith R. Tinsley at Intel

Thursday, December 2, 2010

Slides

Abstract

This talk updates our progress in our Intel-funded project for radio frequency interference (RFI) modeling and mitigation, and future plans for developing RFI-aware cognitive radio algorithms. In some wireless networks, RFI is the limiting factor in communication performance. RFI intensifies with higher reuse of the radio spectrum in general, and the shrinking form factor of the user's computational platform in particular. RFI from external electronic devices, co-existing wireless communication sources, and platform circuitry follow the same non-Gaussian statistics.

First, we derive statistical models of additive RFI in Wi-Fi, Wimax, and other common wireless network topologies, and show a 10-100x decrease in bit error rate when using these models for a single carrier system. Then, we describe more recent advances:

Extensions of our statistical models to incorporate temporal and spatial correlation for space-time communications,
Extensions of communication performance analysis in RFI to include burst errors, delay, throughput and reliability-rate tradeoffs, and
RFI-resistant design of multi-antenna OFDM transmission and reception.

Finally, we discuss plans for RFI-aware cognitive radio algorithms. We have released our RFI modeling and mitigation methods in a free MATLAB toolbox:

http://users.ece.utexas.edu/~bevans/projects/rfi/software/index.html

Questions

Question #1. How do errors in parameter estimates for RFI statistical models affect communication performance of pre-filtering based on those parameter estimates?

Answer #1. The myriad filter for pre-processing has a single parameter whose optimal value is computed as follows:

k(alpha) = sqrt(alpha / (2 - alpha)) y^(1/alpha)

Here, alpha is the exponent in the Symmmetric Alpha Stable distribution where 0 < alpha < 2, and y is the dispersion parameter (analogous to variance).

Here are the values of derivation of k(alpha) with respect to alpha for selected values of alpha:

alpha     k'(alpha)
-----     ---------
0.5       0.7698 sqrt(y) + 0.57735 sqrt(y) ln(y)
1.0       y + y ln(y)
1.5       2.31 y^(3/2) + 1.73205 y^(3/2) ln(y)

For y = 1, k'(alpha) varies from 0.7698 to 2.31 for 0.5 < alpha < 1.5. In fitting RFI data, we have found that alpha > 0.5. Pertubations in k are particularly severe as alpha -> 0 and as alpha -> 2.

Question #2. How does your approach compare to successive interference cancellation?

Answer #2. Interference cancellation, when possible, will provide the most benefit as it reduces the interference itself in the first place. Our approach on mitigating RFI based on statistical models, aims to improve the communication performance by mitigating the residual interference still present at the receiver. The rational behind our approach is:

Interference cancellation may not be possible in many cases, e.g. when interference is from an amalgamation of wireless users using a different standard, electronic devices, etc.
There is always some residual interference present at the receiver. For example, you may be able to cancel interference from other in-cell users in cellular networks, but you may still be affected by out-of-cell users.

The latter is the reason we consider statistical modeling of RFI in different network scenarios (out-of-cell interference in cellular networks and in ad hoc networks with guard zones). Thus our methods can be used in conjunction with other methods to mitigate the residual interference.

Question #3: Lower Kullback-Leibler (KL) divergence does not imply correspondence in tail probabilities. How do we compare the accuracy of the statistical models derived with respect to modeling the tail probabilities?

Answer #3. Accurately modeling tail probabilities is important as the bit error rate performance (or outage probability in networks) depends on the tail probabilities of the interference. We compare the tail probabilities by:

(a) Plotting the tail probability vs. threshold curves for the empirical interference and compare against those obtained from the estimated statistical models. For example:
- Slide 50: Tail probability of platform interference (empirical vs. estimated model tails)
- Slides 42 and 43: Tail probability of interference in various ad hoc and cellular network scenarios defined on Slide 41
- Slides 47 and 48: Tail probability of interference in various femtocell network scenarios defined on Slide 44
(b) In our papers, we used decay rates to compare the correspondence in tail probabilities, where decay rate is the rate at which the tail probabilities are decaying and is defined as:
decay rate at a threshold value = - log (tail probability at that threshold) / threshold
Draft of our journal paper for more information. Decay rate is defined in equation (66). Figures 3-8 compare the decay rate of interference.

Mail comments about this page to bevans@ece.utexas.edu.