Minutes of Meeting, March 24Th, 2004

Minutes of meeting, March 24th, 2004

roll call
- Andy Molisch
- Julien Keignard
- Ulrich Schuster
- Patrick Houghton
- John Goodrich
- Kannan (Singapore)
- Chia-Chin Chong

- Sanjay Mani (Tzero Technologies)


1. admin stuff
- minutes of last meeting adopted
- minutes of all meetings up to Orlando meeting are up on the ieee server

- document on current status of channel model: at ftp://ftp.802wirelessworld.com/15/04/15-04-0024-00-004a-status-report-802-15-4a-channel-modeling-subgroup.ppt

2. Common Data format/date extraction
– No format yet, as there has been no exchange of raw data yet. People have just been exchanging parameters.

- Julien: It is much better to use a database/share raw data and use a common algorithm to extract parameters from that data.

- Andreas: Practically, it is difficult to share data due to proprietary nature, so it is better to at least get some model info.

- due to time restrictions, it was agreed to concentrate on common procedure for data extraction rather than common data format

3. Document for data extraction: will be written up as common “recipe” for data extraction
- write up a document explaining the methods on how to extract these parameters
first draft:


- measurement campaign setup: Andy, John
- large scale parameter extraction: Kannan
- small scale parameters: Uli; Lund will provide document on m-factor estimation
- SV parameters: Chia-Chin Korea

3. Model
- # clusters: extensive discussion on best way to model different number of clusters. Current approach is two-spike pdf. Chia-chin had suggested uniform between min and max. Final decision: change current model to poisson distribution with mean depending on the environment (fixed for one simulation run)
- just use delay-domain only to estimate number of clusters: no angular cluster arrival included
- just use (2) for path loss, no frequency dependency
- path loss: (8) would be nice, but depends on available measurements
- angular spread include if enough measurements are available
3. report Orlando
- status summary, not too many questions
- some channel measurement presentations by CRL -> IEEE server

Detailed minutes taken by

Sanjay Mani

Tzero Technologies

NOTE: I often couldn’t tell who was speaking, and sometimes I’m guessing. Also, I’m not sure how to spell people’s names. Occasionally I use a (#) to indicate the same person.

·  Andreas: Agenda: Adoption of minutes from last meeting, exchange of data, data formats. Report on Orlando. Modelling document and parameterization.

o  Adoption of Minutes – no modifications, adopted.

o  Andreas: Document on IEEE Server:

o  ftp://ftp.802wirelessworld.com/15/04/15-04-0024-00-004a-status-report-802-15-4a-channel-modeling-subgroup.ppt ??

·  Andreas: Common Data Format – No format yet, as there has been no exchange of raw data yet. People have just been exchanging parameters.

o  1: It is much better to use a database/share raw data and use a common algorithm to extract parameters from that data.

o  Andreas: Practically, it is difficult to share data due to proprietary nature, so it is better to at least get some model info.

o  2: Different set-ups mean different types of data, so you can’t do some kind of common analysis.

o  1: With Network analyzer/frequency domain measurements can exchange data, perhaps corrected for system distortion.

o  2: Have to fix bandwidth, antennas, etc. to try to give meaning to raw data.

o  1: 3-5 GHz bandwidth.

o  Andreas: Experience in Metamorph to exchange data, takes too long (1.5 years) to come up with a common data format/script, but our timeline is end of June 2004, so we have to give up this ideal strategy for a pragmatic strategy.

·  Lets focus on the parameters rather than the focusing on coming up with a raw data exchange format and script.

o  (Uly): Nice to share Matlab etc. so that people understand analysis methods.

o  1: How do you extract parameters, pick clusters ? Visual inspection ?

o  Visual inspection seems like the best way.

o  1: Yes, but how exactly do you do this ? Perhaps we should have a document to explain how to do this.

·  Perhaps someone can do a 1st draft of procedures and then move from there.

o  Yes, such a document would be good, but we need volunteers. Who would participate ? John Goodrich, Jacques (sp?), Kanen from Singapore.

o  Process to come up with procedure documents:

o  Extract parameters from Andy’s document.

o  Discuss parameters, and then move forward.

o  Instead of writing a document, we could extract parameters and circulate via email, before directly writing a document.

o  Andreas: I’m more interested in separating small, large, SV parameters.

o  Kanen: I have more experience in time domain, not frequency domain. Are we going to do 2 documents, time domain and freq. domain ?

o  Uly: Just writing up a ½ page document for the process

§  Measurement campaign set-up description: Assigned to Andy, John says he can contribute.

§  Small scale parameter extraction:

§  Large scale parameter extraction: Ananth

§  SV Parameter extraction: Chi-Chin from Korea

o  One of my students looked in the literature at pulling out m parameter.

o  Uly: Please send around paper. Anyone have experience in large scale parameter extraction?

o  I have looked at large scale parameter extraction and circulated a paper.

Go over model document

·  Brought up by Chi-Chin – Should we have 2 discrete values for # of clusters, or should we have some sort of uniform distribution with the parameters would be the upper and lower limits ?

o  Chi-Chin – some environments might have more values of # of clusters than 2, so flexibility would be good.

o  Andreas: Fixed numbers for simplicity in modeling. So for example, for a factory environment, you would have a couple of values. Hard to come up with numbers that are much better for number of clusters – its clearly more realistic, but it may not help the model.

o  Why not fix it to 1 number of clusters ?

o  At the moment we have 2 values for each possible environment. It would be too restrictive to set it to 1 value.

o  Chi-Chin – number of clusters will depend on objects in environment.

o  John: 1 cluster for each plane on the object.

o  Your proposal should work for any environment (factory, indoor, outdoor), meaning different sets of environments, so you should simulate for all environments.

o  Difference between us and ‘3’ is that we have many more environments.

o  You want to simulate your system, so you simulate and get a CDF for probability of errors. The criteria is where does CDF intersect the 90% probability line. You’re not interested in the mean, you’re interested in the worse case. Therefore, the worse case cluster model will determine your performance. It might be the case that if you have a 15% probability of your many cluster option, this will solely determine the performance of your system.

o  John: It is bad to build a system for few clusters and end up in a many cluster environment – its better to shoot for many clusters for a system, 5 is not a bad number.

o  Another idea then would be to look at a worst case and best case number of clusters.

o  John: Want a robust system, so should build for ‘more’ clusters.

o  That’s not the question, the question is do we have 2 delta functions, best and worst, or do we take a uniform distribution between best and worst.

o  The problem is that for a given environment, you tend to have a fixed number of clusters, so you need to look at many different environments to find this best and worst case, and this is unlikely to happen in the timescale we are looking at.

o  Still have to get upper/lower cluster numbers and probability of upper and lower.

o  Uniform model might be less accurate as the cluster probability distribution may not be uniform.

o  Isn’t it enough that someone can change their model for the worst case parameter ?

o  Well, if you want to evaluate the standards proposals you need a fixed test environment.

o  Alternative: Poissen distribution with a mean value for number of clusters.

o  Any reason why Poissen ?

o  Smaller number of parameters than uniform or 2 spike.

o  The reason is that Cost-239 used Poissen – but narrow band outdoor system.

o  On the positive side, Cost-239 did use a ton of measurements.

o  Which is preferred, Poissen or uniform ?

o  Whatever is more peaky is probably better.

o  Perhaps we should go around and ‘vote’.

o  John: How would I identify the number of clusters with Poissen ?

o  Answer: Poissen is characterized by mean, so go out and measure mean and use that as a parameter.

o  John: So say I had a tractor-trailer or some container with 5 reflective sides, creating 5 clusters – can I use that ?

o  Answer: That’s fine for your application, but we need models to test our proposals.

o  John: But if it is something that can be shipped, it could be all sorts of environments.

o  Answer: Could just change the model for a worse case to see how the system performs.

o  John: Right, so with Poissen how would you change the model ?

o  Answer: Use a higher mean.

o  Use model by running a Monte Carlo simulation for a given mean of number of clusters, and then plot CDF of packet error rate and find the required SNR for a reasonable transmission probability, 90%. Then can try again for a different mean. So there is flexibility.

o  Going around seeing who likes which distribution:

§  Jacques: Like Poissen

§  Kanen: Poissen

§  Cha-Chin: Poissen

§  Another for Poissen

§  Robert Shore did a paper looking at number of clusters, an email will be send tomorrow about it.

·  Are we going to count clusters in the angle-delay domain (2D) or in the delay domain. Number of clusters could be quite different.

o  Chi-Chin: What about if you’re using 1D reception – must use delay domain.

o  Since we can always look at 1D with 2D measurements, 1D is probably the way to go.

o  Chi-Chin: 2D would be more useful if we had it, but a lot of people aren’t doing these measurements.

o  Do we need to do ranging with DOA ?

o  Answer: Ranging is a requirement, but no specification on DOA or TOA. I don’t know that there will be multiple antenna systems given low cost goals.

o  Easier to find number of clusters from spatial-temporal measurements.

o  People are going to be doing both, what if the number of clusters varies between the 2 sets of measurements?

o  If we have just a small number of measurements, we might need to just use delay domain.

o  If the majority agree, lets just go with delay domain.

o  Agreed.

Path Loss

·  What is the exact proposal for path loss.

o  Andy: It isn’t necessary to define a frequency dependent path loss. Equation 2 in document should be the equation for the path loss. For distance dependence, Equation 8 is the most advanced equation if we get a lot of measurements, we can model shadowing as a random variable. Otherwise, if fewer measurements, use equation 7.

o  Question: Lets discuss Equation 11. You are going to add this diffuse background radiation parameter to Equation 9, right ? (Yes). Is this an important parameter, I haven’t seen it before.

o  I think it could become more important, not the distribution so much, but the ratio of the power in the diffuse component to the power in the discrete component, because that will determine the performance of a RAKE receiver with a finite # of fingers.

o  OK. With equation 9 and 11, you have 2 small scale components to extract from your model. It is not going to be easy.

o  Yes, from a parameter extraction point of view, it makes it more difficult.

o  So how would you do this ? Identify discrete components, subtract them out, and estimate the rest ?

o  Yes. In narrowband you have a plane wave estimation model, use that, and than subtract out the discrete components and estimate the diffuse parameter from whats left.

o  Is there a paper ?

o  Yes, I will try to send it out. I’ll need to get the author’s agreement, I’ll try.

·  What about the angular stuff in the last section – we may not be measuring it ?

o  I know about a few campaign that are measuring angular spread, if we get the data, great, otherwise its not a big deal.

·  If we do the measurements for 30m, is that enough ?

o  A longer range would be desirable.

o  Again, a question of what can we get. We go with what we’ll get – we’ll state in our model that its only valid out to 30m or whatever.

o  It requires a very high power to do good measurements at distance.

o  Also it is difficult to put a cable between transmitter and receiver.

o  I’m targeting 30 m.

o  That sounds reasonable. Its nice to have a general model, but must keep in mind timescale.

o  30m would be quite good.

·  What about that email from Kai ?

o  If you assume a single exponential power delay profile, and you vary the delay, then the power in the first cluster will scale obviously for fixed total power, so I didn’t see the point.