The Cost of Interrupted Work: More Speed and Stress

This article stated this statistic: "And after each distraction, it takes on average 23 minutes and 15 seconds to truly get refocused on the task at hand."  And cited this paper as the reference.  That is why I'm here.

So, after reading the paper I can confirm that the 23 minutes comment is totally wrong.  I wonder if the article writer actually read the paper.  Instead, the paper's results showed that there was not a difference in time taken to complete a task with interruptions but the user's stress level rose significantly (0.05).


Mark, Gloria, Daniela Gudith, and Ulrich Klocke. "The cost of interrupted work: more speed and stress." Proceedings of the SIGCHI conference on Human Factors in Computing Systems. ACM, 2008.

Best Linear Approximation

This is a nice 30 page overview of Linear System Identification in a Nonlinear Setting.

• Figure 7.  They did some interesting validation of the BLA that reminds me of the JMP analysis I used to do a lot of.  This maybe a nice enhancement to our psuedo-impulse approach. (maybe psuedo-linear is a better name?)
• "The major advantage of the random-phase multisine is that it still has (asymptotically for sufficiently large N) all the nice properties of Gaussian noise, while it also has the advantages of a deterministic signal: the amplitude spectrum does not show dips at the excited frequencies [see Figure 10(f)] as the two other signals do [see Figure 10(d) and (e)]. At those dips, the measurements are very sensitive to all nonlinear distortions and disturbing noise."
• RJA: This would be interesting to investigate to see if this has an advantage over the PRBS patterns we use.
• "The major goal [is to discuss the] 1) detection and characterization of nonlinear distortions, 2) extending the linear framework to include the effect of nonlinear distortions, and 3) quantifying the potential gain by switching from a linear to a nonlinear identification framework."
• 
  
  
  
• RJA: I should try creating one of these random-phase multisine signals
• "the excitation signals during the experiments should match as well as possible the signals that will be applied later on to the model "
• I should check this out again: [32] H. K. Wong, J. Schoukens, and K. R. Godfrey, “Analysis of best linear approximation of a Wiener-Hammerstein system for arbitrary amplitude distributions,” IEEE Trans. Instrum. Meas., vol. 61, no. 3, pp. 645–654, 2012
• "This allows well-selected pseudorandom binary excitations to be used in many practical applications to measure [32].
• "All the results in this article can be reproduced using publicly available Matlab toolboxes. The motivational example was produced using the System Identification toolbox of Matlab (Mathworks). Alternatively, the freely available frequency-domain identification toolbox FDIDENT could be used to obtain similar results (http://home.mit.bme.hu/~kollar/fdident/). This toolbox also includes the tools to design the random-phase multisines and to perform the nonparametric nonlinear analysis. In [4], all the procedures that are presented in this article are discussed in full detail, and the related Matlab software can be freely downloaded from booksupport.wiley.com."BLA[32

J. Schoukens, M. Vaes and R. Pintelon, "Linear System Identification in a Nonlinear Setting: Nonparametric Analysis of the Nonlinear Distortions and Their Impact on the Best Linear Approximation," in IEEE Control Systems Magazine, vol. 36, no. 3, pp. 38-69, June 2016.

polynomial chaos resources

"Fast Numerical Methods for Stochastic Computations: A Review" by Dongbin Xiu∗
http://www.ece.uvic.ca/~bctill/papers/numacoust/Xiu_2008.pdf

Prof Xiu wrote the book on the topic


"Polynomial Chaos: A Tutorial and Critique from a Statisticianís Perspective" by Anthony OíHagan
http://tonyohagan.co.uk/academic/pdf/Polynomial-chaos.pdf

Derivatives of TDR

A quick googling showed that for signal integrity applications, I couldn't find any mention of using the derivative of the TDR for any type of analysis but it turned out that it is common for water soil studies!  Here are a few links that I don't want to forget:

"A Comparison of Second-Order Derivative Based Models for Time Domain Reflectometry Waveform Analysis" in the alliance of crop, soil and environmental science societies
https://dl.sciencesocieties.org/publications/vzj/abstracts/16/7/vzj2017.01.0014?access=0&view=pdf


A nice picture of the TDR and 1st Derivative:
https://www.researchgate.net/figure/Plot-of-time-domain-reflectometer-TDR-waveform-and-first-derivative-as-seen-on-a-TDR_fig1_266878451

On the effective measurement frequency of time domain reflectometry in dispersive and nonconductive dielectric materials
https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2004WR003816


I think that including these papers would give a nice multi-disciplinary flair to a future paper.

A novel simplified four-port scattering parameter model for design of four-pair twisted-pair cabling systems for local area networks

Reaction: This paper has many similar elements to my work.  Their primary objective is to create a fast cascading method for 4 port S-parameters, secondary consideration is the intuitive understand of the combination of S-parameters but they don't really develop this idea.
They only formulate the combination of 2 4-port S-parameters and require a repetitive application of the their method for more S-parameters.  This limits the insight to only a system of 2 S-parameters.
Their method primarily discards any terms with 2 or more xtalk terms and is therefore not applicable to ISI studies of 2-port systems like mine.  I believe that their method would yield near exact results of the 2-port S-parameter cascading.  

Their work is very similar to my xtalk decomposition Yun and I developed at Intel.  If I ever go back to that topic, this paper will certainly be there too.


Huang, C-WP, Charles E. Smith, Atef Z. Elsherbeni, and B. H. Hammond. "A novel simplified four-port scattering parameter model for design of four-pair twisted-pair cabling systems for local area networks." IEEE Transactions on Microwave Theory and Techniques 48, no. 5 (2000): 815-821.

High-Speed Channel Modeling With Machine Learning Methods for Signal Integrity Analysis

Overview: A nice overview of machine learning applied to a simplistic interface.  They don't share a lot of details so it would be difficult to replicate.

Impression:  It looks like they are applying machine learning to a simplistic problem of simple transmission line properties along with jitter and tx/rx equalization.  They are not specific about what the input variables are.  They seem to choose examples which favor their approach.  They do show a good comparison between a few different machine learning approaches and variants.  Overall, this has potential but not super impressed.  It would be an interesting exercise to try and replicate their work (although they don't really give enough info to do this).

Summary:

• Give a bunch of SVR (support vector regression) math.
• "To empower SVR to handle nonlinearity, a natural idea is to map the input vectors {xi} to a high-dimensional feature space through a nonlinear mapping, denoted as Φ({xi}), and then feed these mapped vectors Φ({xi}) into the linear model"
• Looks like they used Python
• They are very light on details and their many layer make it difficult to figure out what they did exactly.
• Figures 5 and 6 show a sweep of tx pre-emphasis and the resulting eye height.  This is a very strange result.  Why would their eye height change in a non-smooth manner unless their pre-emphasis are not a linear increase in strength but some set of pre/post strength like used in PCIe3.  I wonder if they are picking examples which show their method as best.

T. Lu, J. Sun, K. Wu and Z. Yang, "High-Speed Channel Modeling With Machine Learning Methods for Signal Integrity Analysis," in IEEE Transactions on Electromagnetic Compatibility, vol. 60, no. 6, pp. 1957-1964, Dec. 2018.

An Efficient High-Speed Channel Modeling Method Based on Optimized Design-of-Experiment (DoE) for Artificial Neural Network Training

Main Idea: Instead of using just random training cases to train the neural network, to use Design of Experiments to get better coverage with less cases.

Not super impressed but after a read of the abstract+conclusion and scan of the figures, I believe what they are doing is simply taking a MTL (multiple transmission line) and using an artificial neural net (ANN) to predict the MTL RLGC from input parameters.  This looks like a nice toy application but not immediately useful.


H. Kim, C. Sui, K. Cai, B. Sen and J. Fan, "An Efficient High-Speed Channel Modeling Method Based on Optimized Design-of-Experiment (DoE) for Artificial Neural Network Training," in IEEE Transactions on Electromagnetic Compatibility, vol. 60, no. 6, pp. 1648-1654, Dec. 2018.

Machine Learning for the Performance Assessment of High-Speed Links

Comparison of SVM (support vector machines) to PC (polynomial chaos) for surrogate model creation of high-speed interconnects with largely varying and/or high uncertain design parameters.

• Nice summary of Polynomial Chaos expansion.
• SVM regression (also known as epsilon-regression)
• They reference statistical learning literature.  This appears to be a classic, take a well known method from another domain and apply it to our problems.
• I stopped reading after a page or two of math.  I would be interested to return to this study in the future.

R. Trinchero, P. Manfredi, I. S. Stievano and F. G. Canavero, "Machine Learning for the Performance Assessment of High-Speed Links," in IEEE Transactions on Electromagnetic Compatibility, vol. 60, no. 6, pp. 1627-1634, Dec. 2018.

Design of Multilevel Signals for Identifying the Best Linear Approximation of Nonlinear Systems

[H. K. Wong, J. Schoukens and K. R. Godfrey, "Design of Multilevel Signals for Identifying the Best Linear Approximation of Nonlinear Systems," in IEEE Transactions on Instrumentation and Measurement, vol. 62, no. 2, pp. 519-524, Feb. 2013.]

BLA: Best Linear Approximation

This looks like an interesting paper but they have a different requirement on the input sequence.  Where I am most interested in an good autocorrelation they want all moments of the signal to be zero.

The BLA depends on the power and the amplitude distributions of the inputs {u}. As noted earlier, the BLA when using a Gaussian input sequence has been well studied [2], [6]. The objective in this paper will be to design multilevel sequences with moments as close as possible to those of a zero-mean Gaussian sequence u∼N(0,σ2), for which the mth moment Mm≡E[um]=(m−1)!! for even m and zero for odd m. Here, (m−1)!! is the double factorial of (m−1) given by (m−1)(m−3)(m−5)⋯1. The closer the moments of an arbitrary sequence match the moments of N(0,σ2), the closer the estimated BLA will be to that estimated using a Gaussian input sequence.