1. Accurate WiFi Packet Delivery Rate Estimation and Applications Owais Khan and Lili Qiu. The University of Texas at Austin 1 Infocom 2016, San Francisco.

2. 2 Frequency Diversity Wireless channel are frequency selective Bandwidth is increasing with new standards BW (MHz) Channel Subcarriers SNR (dB) ~22dBs 20 40160 1825 Wider Channels experience more frequency diversity

3. 3 Challenges Frequency diversity results in bursty errors – Lower SNR subcarriers have more errors Delivery Rate hard to predict for bursty errors even with Channel State Information (CSI) Existing techniques do not perform well in bursty errors

4. Limitation of Average SNR Avg. SNR is widely used for delivery rate estimation SNR is not a good predictor in Freq. Selective Channels Delivery Rate Ref. Halperin10 12 dB Gap 4

5. Effective SNR Effective SNR shows better results than Avg. SNR Approach 1.Map the SNR per subcarrier to BER 2.Map BER eff,k back to Effective SNR 3.Use Effective SNR to select the appropriate rate How Accurate is Effective SNR? 5

6. Effective SNR Accuracy Effective SNR is also not very accurate Scatter plot of Estimated delivery rate vs. gnd. truth 95% 39% Gap of 56% 6

7. Error Burstiness across Frame code rate: 1/2 802.11 interleaver does not uniformly distribute bits 7 WiFi interleaver has a skewed distribution Error still bursty in WiFi interleaver 65% Delivery rate estimation must incorporate error burstiness

8. Contributions Develop two new techniques to estimate delivery rate in bursty errors Propose a new Interleaver to reduce burstiness 8

9. Goal – To accurately predict delivery ratio using CSI information Problem Formulation Options 1.Analytical modeling Intractable for freq. selective channels. 2.Simulate online Prohibitively expensive 3.Lookup table based approach Pre compute delivery rate for error patterns Error patterns capture the burstiness 4.Machine learning based approach Use supervised learning to estimate delivery rate 9 Our Approach

10. 10 Lookup Table: Idea Decoding success depend on error pattern to FEC Decoder – Same errors have different outcome depending on location Receive Chain De-ModulationDe-InterleaveFEC Decode Data Analog Signal 0.5+1.2j, -2-0.8j,… 1, 0, 1, 1, 0, … Idea: “Calculate the error probability of all possible patterns offline and use them to estimate delivery rate for given CSI”

11. Run Viterbi decoding offline and build a lookup table Use error pattern as input Lookup Table: Approach Independent of wireless channel or hardware Using an error pattern for frame prohibitively expensive 1 1 0 0 0 0 0 1 0 0 1 1 0 0 0 0 0 0 1 1 0 0 0 0 0 1 1 1 1 1 0 0 0 0 0 0 0.1 0.9 … … 11

12. Errors in frame decouple after a certain period Lookup Table: Observation Error seq. 1Error seq. 2 The de-coupling length is dependent on code rate Code RateWindow Size 1/275 bits 2/350 bits 3/450 bits 5/640 bits 2^40 is still a very large number! Independent! 12

13. We find upper and lower error thresholds – Errors less than lowerThresh, decoding always successful – Errors greater than upperThresh, decoding always fails Can we reduce size further? Code Rate LowerThreshUpperThresh 1/2510 2/335 3/424 5/624 Complexity reduces to nChoosek 13

14. To Summarize 1.Each code rate has a fixed window size 2.The table is bounded by number of errors in a window Implement the table as a hash map – The error sequence is converted to integer and used as key Actual Lookup Table 1 1 0 0 0 0 0 1 0 0 1 1 0 0 0 0 0 0 1 1 0 0 0 0 0 1 1 1 1 1 0 0 0 0 0 0 0 0.75 … … … 3203 2907 14 How to use these tables to estimate delivery rate?

15. Delivery Rate Estimation Calculate BER per subcarrier from SNRGenerate Error Patterns from BERLookup Error Prob. for error patternLookup next pattern in sequenceTake product of all Error Probs.Repeat for multiple error sequences 15 Error Pattern Generation

16. Sliding window operation b1b1 b3b3 b4b4 b2b2 b5b5 b6b6 b8b8 b7b7 b9b9 b 10 b 11 b 12 b 13 b0b0 b 14 b 15 1100010001010000 p1p1 p2p2 p3p3 P e,1 = p 1 x p 2 x p 3 Example: Window Size = 4 Delivery Rate is the average of Prob. of Error of N seqs 16

17. Machine Learning Motivation – Avoids the time and space complexity of lookup tables – Machine learning can provide faster online solution Machine learning algorithm – We chose Neural Networks Reason – Supports non-linear continuous functions – Appropriate for delivery ratio 17

18. Feature Set – Use Bit error Rate per bit Advantage – Easy to obtain from CSI information – Allows de-coupling from the interleaver – Feature size is limited by number of bits in OFDM symbol 18

19. Neural Network: Setup 19 Input Layer Hidden Layers Output Layer Multi Layer Perceptron Bit Error Rate 2 Layers 20 Neurons per layer Neuron Delivery Rate

20. Neural Network Operation Neural Network Network weights Training Intel Channel Traces TGn Channel models Bit-Error-Rate Delivery Ratio Input Output Testing Trained Network Delivery Rate Bit-Error-Rate Input Output 20

21. Contributions Develop two new techniques to estimate delivery rate Propose a new Interleaver 21

22. Standard Interleaver Spreads adjacent bits by four subcarriers 2728…39 4041…52 1415…26 12…13 1142740 2152841 … 132652 Subcarriers 4 spaces apart can experience similar fate 39 12345678…49505251 22

23. Proposed Interleaver Key Ideas 1.Maximize the separation between likely corrupted bits 1.Use SNR information to identify more error prone bits Interleaving pattern dependent on SNR 23

24. Proposed Interleaver 10 Subcarriers s1s2s3s4 s5s6s7s8s0 s9 Sorted based on SNR 1234567890 Maximize separation between bad locations s0 s1 maximum separation next ofdm symbol 24

25. Proposed Interleaver 1234567890 s0 s1 s2 s3 s6 s4s5s7s9s8 b1b1 b3b3 b4b4 b2b2 b5b5 b6b6 b8b8 b7b7 b9b9 b0b0 maximum separation 25 s1s2s3s4 s5s6s7s8s0 s9 10 Subcarriers Sorted based on SNR Maximize separation between bad locations Interleaver maximizes the separation between bad bits

26. 26 Evaluation and Results

27. Evaluation Methodology Trace driven simulation to evaluate performance Traces collected using Intel 5300 WLAN card Trace Features – Static and mobile traces – Both 20Mhz and 40Mhz bandwidths Traces generated using IEEE-TGn Channel models – Traces simulate a range of indoor and outdoor environments – Include both 20Mhz and 40Mhz bandwidths 27

28. Evaluated Schemes SchemesDescription EffSNR Average BER is used to estimate effective SNR Lookup Our lookup table based approach ML Trained neural network is used to estimate delivery ratio 28

29. Delivery Rate Accuracy (40MHz) EffSNR Wifi: 11% Our: 27% Lookup Wifi: 4.5% Our: 3% ML Wifi: 6% Our: 3% 29 Lookup and ML show improvement over Effective SNR

30. Interleaver Performance 20 MHz40 MHz 17% impv. 22% impv. 30 Proposed interleaver outperforms 802.11 interleaver

31. Rate Adaptation 20 MHz40 MHz 31 Lookup Wifi: up to 62-75% Our: up to 72-79% ML Wifi: up to 65-75% Our: up to 72-79% Significant throughput benefit at transition regions

32. Energy Savings Transmitter 20 MHz40 MHz 32 Lookup Wifi: up to 26-35% Our: up to 29-36% ML Wifi: up to 27-35% Our: up to 29-36% Significant Energy savings at transition regions

33. Summary Identify the reasons for poor performance Propose Lookup and ML based approach to estimate delivery ratio – Reduce delivery rate error by 60% for 802.11 interleaver and 88% for proposed interleaver Future Work – Explore other applications of machine learning in wireless network management 33

34. THANK YOU! 34