9999 parameter optimizing

Upload: hari-pramono

Post on 05-Apr-2018

238 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/2/2019 9999 Parameter Optimizing

    1/15

    INTERNAL INFORMATION

    REPORT 1 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    AN AUTOMATIC METHOD FOR OPTIMISING

    PARAMETERS OF ALGORITHM 9999

    CONTENTS:

    1 INTRODUCTION.................................................................................................................................2

    2 BACKGROUND...................................................................................................................................2

    3 METHODOLOGY ...............................................................................................................................2

    3.1 BASIC ASSUMPTIONS .....................................................................................................................2

    3.2 THEORY.............................................................................................................................................3

    3.3 FINAL DETERMINATION OF A0 AND A1......................................................................................5

    4 LIMITATIONS.....................................................................................................................................5

    4.1 NOT ALL PARAMETERS OPTIMISED ............................................................................................5

    4.2 NO KNIFE-EDGE DIFFRACTION.....................................................................................................6

    4.3 NO SPHERICAL EARTH LOSS.........................................................................................................6

    4.4 NOT CORRECT EFFECTIVE BASE ANTENNA HEIGHT...............................................................6

    5 POSSIBLE SOURCES OF ERRORS ..................................................................................................7

    5.1 BAD STATISTICS FOR A CLUTTER................................................................................................7

    5.2 BAD QUALITY OF MAP DATA........................................................................................................7

    5.3 TOO MANY DATA POINTS..............................................................................................................9

    5.4 NUMERICAL PROBLEMS WITH DETERMINING A1....................................................................9

    6 RESULTS IN ROME..........................................................................................................................10

    6.1 URBAN GLOBAL MODEL..............................................................................................................10

    6.2 SUBURBAN GLOBAL MODEL.......................................................................................................11

    7 CONCLUSIONS .................................................................................................................................12

    8 REFERENCES....................................................................................................................................13

    APPENDIX: DESCRIPTION OF HOW TO USE THE MATLAB PROTOTYPE

  • 8/2/2019 9999 Parameter Optimizing

    2/15

    INTERNAL INFORMATION

    REPORT 2 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    1 INTRODUCTION

    A method for optimising steering parameters in algorithm 9999 has been

    developed within the project GLENN [1]. The tool used for this optimisation hasnot yet been implemented into EET and consists at present of a prototype in

    Matlab. Using this prototype on measurements shows that the method works

    well and that it is possible to achieve qualitative results in a short amount of

    time.

    This report describes the methodology of optimisation and also discusses some

    limitations with the method. Also some possible error sources and ways to

    handle them are discussed. Then results of an optimisation made onmeasurements performed in Rome are presented.

    2 BACKGROUND

    The process of optimising the steering parameters in algorithm 9999 has up until

    now been time consuming and tricky. The method has so far been to change one

    variable at the time in small steps and then do a survey analysis for each settings.

    There are several number of iterations to perform, in order to find the smallest

    RMS error and standard deviations. Thus the normal way of optimising 9999has been to simply adjust the clutter values in order to minimise the mean errors,

    A0 to A4 has normally been set to their default values.

    It is quite obvious that the solution of this problem is to develop a method for

    automatic parameter tuning.

    3 METHODOLOGY

    3.1 BASIC ASSUMPTIONS

    At present, it is difficult to get data directly from the 9999 algorithm. This can

    eventually be changed in a future, due to implementation of the EET API. It is a

    possibility that this interface can be used for direct communication with

    implemented models, such as algorithm 9999, e.g. regarding values of model

    variables. But so far one has to use other methods for extracting model data and

    in this case output files from the Survey Analysis Tool in EET are used for

    obtaining necessary data.

  • 8/2/2019 9999 Parameter Optimizing

    3/15

    INTERNAL INFORMATION

    REPORT 3 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    This means that it is essential to make some simplifications. Neither knife-edge

    diffraction nor spherical earth loss is taken into account. Further only the

    parameters A0 and A1 are optimised, due to some numerical trickiness, all

    parameters are in fact affected by each other in some degree. A2 and A3 are

    assumed to be known (nowadays A4 is always set to A1). The optimisation

    method used is the least square method.

    3.2 THEORY

    Considering no knife edge contribution and no spherical earth loss, the path loss

    according to 9999 is [ ]2 :

    [ ]L mk mobil A A d A H A d Hp eff eff = + + + + 0 1 2 3log log log log

    ( )[ ] ( ) +3 2 11 752

    . log . H g Fm [ ]dB

    where [ ]mk mobil : value of land usage code at mobile [ ]dB

    d : distance from base antenna to mobile [ ]km

    Heff : effective height of base antenna [ ]m

    Hm : height of mobile antenna [ ]m

    ( ) ( )g F F F= 44 4 782

    .49 log . log

    where F : frequency [ ]MHz

    A0, A1, A2 and A3 are prediction parameters. Let [ ]A A mk mobil0 0*

    = + ; inthis case, A0* and A1 are going to be optimised, A2 and A3 are assumed to be

    known. The optimisation is then performed for one clutter at a time.

    We use the most common method of them all to optimise the path loss formula:

    the least square method. We intend to minimise the sum of the difference

    between predicted values and measured data. From the Survey Analysis files, we

    get information about measured signal strength, SSmeas [ ]dBm . The path loss

    Lmeas then can be obtained by using a simple link budget formula,

    L EIRP SS L Gmeas meas fm m= + [ ]dB

  • 8/2/2019 9999 Parameter Optimizing

    4/15

    INTERNAL INFORMATION

    REPORT 4 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    where EIRP : emitted power [ ]dBm

    L fm : mobile antenna cable loss [ ]dB

    Gm : mobile antenna gain [ ]dB

    EIRP, Lfm and Gm are supposed to be known from the actual measurement

    campaign.

    The error function is then as follows:

    ( ) ( )[ ]E A AN

    L A A Lp i meas ii

    N

    0 11

    0 12

    1

    , ,,*

    ,= =

    where N : number of measured points

    L A A d Cp i i i,* log= + +0 1 , i = 1, 2, ... N

    ( )[ ] ( )C A H A d H H g Fi eff i i eff i m= + +2 3 3 2 11 752

    log log log . log ., ,

    A2 and A3 are put to recommended values [ ]3 , Hm and F are known from theactual measurements. d i , Heff i, and SSmeas i, are extracted from the Survey

    Analysis output files and Lmeas i, for each measured point obtained from SSmeas i, .

    For minimising ( )E A A0 1*, , the function is differentiated partially with respectto A0* and A1. Let B di i= log . There will be N equations to be solved:

    i A A B C L

    i A A B C L

    i N A A B C L

    meas

    meas

    N N meas N

    = + + =

    = + + =

    = + + =

    1 0 1

    2 0 1

    0 1

    1 1 1

    2 2 2

    :

    :

    :

    *

    ,

    *

    ,

    *

    ,

    M M M

    This overdetermined equation system (2 unknown, N equations) can also be

    written as,

    1

    1

    1

    0

    1

    1

    2

    1 1

    2 2

    B

    B

    B

    A

    A

    L C

    L C

    L CN

    meas

    meas

    meas N N

    M M M

    =

    *

    ,

    ,

    ,

    or W a Y =

  • 8/2/2019 9999 Parameter Optimizing

    5/15

    INTERNAL INFORMATION

    REPORT 5 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    The desired vector a AA

    =

    0

    1

    *

    is then obtained from,

    [ ]a W W W YT T= 1

    .

    3.3 FINAL DETERMINATION OF A0 AND A1

    Since the optimisation process is made per clutter, we get a set of different

    values of A0* and A1. The overall value of A1 is then determined from the

    formula,

    An

    n Atot

    i i

    i

    N

    11

    1

    1

    ==

    where N : number of clutters used in the optimisation

    n i : number of samples for clutter number i

    A i1 : value of A1 for clutter number I

    n n n ntot N= + + +1 2 . .. : total number of samples

    Regarding [ ]A A mk mobil0 0* = + , it has been shown [1] that the best result isobtained by simply setting A0 to the default value, 36.2. The clutter codes,

    [ ]mk mobil , are then optimised in the conventional way, by minimising the mean

    errors between the prediction and the measurements [4].

    4 LIMITATIONS

    4.1 NOT ALL PARAMETERS OPTIMISED

    At the present, only A0* and A1 (i.e. only A1) are optimised. There needs to be

    investigated whether the result can be improved with a complete optimisation.

    Due to some numerical problems, it is hard to optimise all four parameters, A0*

    to A3 at a time. One idea may be a two step optimisation. In the first step, A0*

    and A1 are optimised, while A2 and A3 are set to their default values. In the next

    step, A2 and A3 are optimised, while A0* and A1 are set to the values obtained

    from the first step.

  • 8/2/2019 9999 Parameter Optimizing

    6/15

    INTERNAL INFORMATION

    REPORT 6 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    4.2 NO KNIFE-EDGE DIFFRACTION

    Since knife-edge diffraction is ignored, optimisation in areas with a lot of knife-

    edges may give large errors. This may in a future be avoided by getting

    information about knife-edges via the EET API. But so far there is required

    measurements from areas with no or a few knife-edges for a good optimisation

    result.

    4.3 NO SPHERICAL EARTH LOSS

    So far no spherical earth loss is included and thus there is a risk that large errorswill occur when predicting long distances. For using the optimisation algorithm,

    it is suggested that measured data are taken at a maximum distance of 10 km

    from the site.

    4.4 NOT CORRECT EFFECTIVE BASE ANTENNA HEIGHT

    The Heff values, obtained from the Survey Analysis output files, are not the

    actual effective antenna heights, according to the 9999 algorithm. They are

    approximated by the height differences between site position and mobile

    position, [ ] [ ]H h transmitter h mobileeff . But in the future, when the EET APIwill be available, it will be possible to get all 9999 parameters directly. Thus the

    true Heff values will be obtained and used in the optimisation method.

  • 8/2/2019 9999 Parameter Optimizing

    7/15

    INTERNAL INFORMATION

    REPORT 7 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    5 POSSIBLE SOURCES OF ERRORS

    5.1 BAD STATISTICS FOR A CLUTTER

    If there are just a few measured points for a certain clutter type, the result of the

    optimisation will be misleading, see figure below.

    0.3 0.31 0.32 0.33 0.34 0.3542

    44

    46

    48

    50

    52

    54

    56

    log d

    M-C

    Figure 1. Example of how bad statistics for a clutter type can affect the result.

    Even if the A1 value from such a clutter type will not affect the overall value of

    A1 to a great extent, it should anyway be sorted out. When performing an

    optimisation, it ought to be possible to exclude clutters with a few measured

    points. In this case, it is also of course very important to judge which clutter

    types to be removed from optimisation procedure.

    5.2 BAD QUALITY OF MAP DATA

    Bad map data quality can also affect the result of the optimisation. In the figure

    below, which is taken from measurements performed in Jakarta [5], we see anexample of this. Points that should belong to another clutter type cause

    preposterous values of A0* and A1.

  • 8/2/2019 9999 Parameter Optimizing

    8/15

    INTERNAL INFORMATION

    REPORT 8 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 150

    60

    70

    80

    90

    100

    110

    log d

    M-C

    Figure 2. Example of how bad map data for a clutter type can affect the result.

    After manual removal of these deviating points, we after a new optimisation get

    a reasonable result, see figure below.

    0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 165

    70

    75

    80

    85

    90

    95

    100

    105

    110

    log d

    M-C

    Figure 3. Result of optimisation after editing of points.

    It is essential to be able to identify points belonging to another clutter and also

    have the possibility to remove these points. If the result, together with the

    measured data can be displayed, it is quite simple to pick out the points that

    caused the misleading result. Then there is required some sort of editor for

    removing deviating points.

  • 8/2/2019 9999 Parameter Optimizing

    9/15

    INTERNAL INFORMATION

    REPORT 9 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    5.3 TOO MANY DATA POINTS

    The maximum number of survey files that can be loaded simultaneously into the

    Survey Analysis Tool is 50. If there are more survey files than that, some of the

    files need to be excluded, which will affect the result. A general experience is

    that the value of the slope of a line, i.e. A1, is very sensitive to the choice of data

    points. A suggestion in this case may be that optimisations are made for a set of

    survey files at a time, and then the overall result is determined by weighting

    together results from the different optimisations. This is similar to the method

    described in section 2.3.

    5.4 NUMERICAL PROBLEMS WITH DETERMINING A1

    Figure 4. The standard deviation function with respect to A1.

    The error function, i.e. the squared standard deviation, is minimised with regards

    to A0* and A1 in order to obtain the optimal values of A0* and A1. However it

    can be shown that the standard deviation as a function of A1 has a very shallowand not very well determined minimum, which can cause misleading A1 values.

    E.g. A1 < 20 will lead to a path loss falling below that of free space loss in some

    points. Experience shows that 25 1 40 A , so there is a suggestion that it willbe a check in the algorithm so that A1 will not fall out of this interval.

  • 8/2/2019 9999 Parameter Optimizing

    10/15

    INTERNAL INFORMATION

    REPORT 10 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    6 RESULTS IN ROME

    The optimisation method has been tested on 1800 MHz measurements

    performed in Rome. Two sets of parameters should be determined, one for a so

    called Urban Global model and one for a Suburban Global model. The results are

    shown below.

    6.1 URBAN GLOBAL MODEL

    For optimisation, there were enough number of samples for five clutter codes, as

    shown in the table below:

    Table 1. Values of A1 for different clutter types.

    Clutter type: Number of points, ni A1i

    urban 2744 24.3684

    suburban 1733 23.0167

    park land 154 20.4178

    major roads and railways 63 39.7260

    industrial (edited) 189 32.1200

    The overall value of A1 was then set to,

    An

    n Atot

    i i

    i

    11

    1 24 2

    1

    5

    = ==

    .

    The total number of points used for the optimisation, n tot , was in this case 4883.

    A0 was set to the default value, 36.2, and the clutter factors were determined to

    be:

    Table 2. Clutter values and standard deviation errors for different clutter types.

    Clutter type: Clutter value (dB): Standard deviation (dB):

    urban 18.3 10.4

    suburban 14.0 10.2

    industrial 8.6 8.3

    open areas (4.1) 4.1

    agricultural land 7.8 9.7

    park land 17.7 9.5

    major roads and railways (16.2) 10.0

    construction sites (18.2) 1.7

    sport facilities 12.8 11.6

  • 8/2/2019 9999 Parameter Optimizing

    11/15

    INTERNAL INFORMATION

    REPORT 11 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    Since there was bad statistics (a few measured points) for clutter factors

    open areas, major roads and railways and construction sites, those clutter

    values should not be trusted and have to be estimated.

    The overall RMS error turned in this case out to be 10.3 dB.

    6.2 SUBURBAN GLOBAL MODEL

    For optimisation, there were enough number of samples for seven clutter codes,

    as shown in the table below:

    Table 3. Values of A1 for different clutter types.

    Clutter type: Number of points, ni A1i

    urban 1763 22.8943

    suburban 2701 21.4882

    sport facilities 211 19.0613

    port areas 73 25.6415

    park land 95 30.6169

    industrial 373 28.0733

    agricultural land 1778 24.5621

    The overall value of A1 was then set to,

    An

    n Atot

    i i

    i

    11

    1 231

    1

    7

    = ==

    .

    The total number of points used for the optimisation, n tot , was in this case 6994.

    A0 was set to the default value, 36.2, and the clutter factors were determined to

    be:

  • 8/2/2019 9999 Parameter Optimizing

    12/15

    INTERNAL INFORMATION

    REPORT 12 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    Table 4. Clutter values and standard deviation errors for different clutter types.

    Clutter type: Clutter value (dB): Standard deviation (dB):

    urban 17.8 9.7

    suburban 12.6 10.3

    industrial 10.0 9.2

    open areas (8.1) 6.6

    plantations 0.8 11.6

    agricultural land 8.3 9.9

    park land (7.0) 9.1

    airports (21.3) 1.3port areas (7.8) 8.5

    major roads and railways (4.6) 10.9

    construction sites (1.3) 8.2

    sport facilities 14.1 9.4

    mineral extraction sites (15.0) 2.0

    Since there was bad statistics (a few measured points) for clutter factors

    open areas, park land, airports, port areas, major roads and railways,

    construction sites and mineral extractionsites, those clutter values should not

    be trusted and have to be estimated.

    The overall RMS error turned in this case out to be 10.1 dB.

    Worth noting in this case is also that the work of optimisation, using the existing

    prototype, took slightly more than one day to perform.

    7 CONCLUSIONS

    It is clear that with the optimisation method described, it is possible to achieve

    good results in a small amount of time. But if this method will be implemented asa tool in EET, there are still some problems regarding technical solutions, that

    need to be investigated. The most important thing is anyway that the user of this

    tool must be aware of what she is doing, use her sense and not just blindly trust

    the obtained result. This optimisation tool will not simplify the work, only make

    it faster.

    In order to avoid misleading results of the optimisation of 9999 parameters, e.g.

    very low A1 values, some kind of decision support should be included in the

    optimisation tool.

  • 8/2/2019 9999 Parameter Optimizing

    13/15

    INTERNAL INFORMATION

    REPORT 13 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    8 REFERENCES

    [1] Molander/Thiessen, Glenn Project Technical Report, Doc. No. 1/0363-FCP 103 651 Uen, Rev A, 1995.

    [2] Melin, L., Functional Specification: EPA - Ericsson PropagationAlgorithm, Doc. No. 155 17-CNL 113 156 Uen, Rev A, 1994.

    [3] Setterlind, C. J., 1800 MHz Parameters For Algorithm 9999, Doc. No.LT/SN-95:375, Rev A, 1995.

    [4] Gullin, J. Guideline For Optimising The 9999 Parameters, Doc. No.LV/R-96:171, Rev A, 1996.

    [5] Thomssen/Thiessen, RF Measurements In Jakarta, Indonesia, Doc. No.ERA/LN/IDN-96:-0089 Uen, Rev A, 1996.

  • 8/2/2019 9999 Parameter Optimizing

    14/15

    INTERNAL INFORMATION

    REPORT 14 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    APPENDIX:DESCRIPTION OF HOW TO USE THE MATLAB PROTOTYPE

    1. Use an output file from the Survey Analysis Tool in EET. In this example, the name is surv_a.txt.

    Note that the coordinates in the file must be given in Decimal Lat-Long format!

    2. Enter MS-DOS and run the program glenn_in.exe.

    E.g. glenn_in surv_a.txt

    3. Now we have data separated in different clutter codes, e.g. open, vegetati, medium, light, dense.

    These have the form

    M M M M

    M M M M

    d SS h SSmeas eff pred

    where

    d: distance [ ]mSSmeas : measured signal strength [ ]dBmh

    eff: estimated effective antenna height [ ]mSSpred : predicted signal strength [ ]dBm according to 9999 algorithm (just for comparison)

    4. Enter Matlab. Each file now can be loaded:

    E.g. load medium

    Then AA = medium;

    5. The matrix AA contains the vector heff, which needs to be cleaned from values less than or equal to 0.

    This is made with the command (program) rensa.

    6. Run the command (program) o9999, which gives the parameters A0 and A1.

  • 8/2/2019 9999 Parameter Optimizing

    15/15

    INTERNAL INFORMATION

    REPORT 15 (15)Uppgjord - Prepared Tfn - Telephone Datum - Date Rev Dokumentnr - Document No.

    LVR/DT Maria Thiessen 40:45228 1997-06-19 A LVR/D-97:196Godknd - Approved Kontr - Checked Tillhr/Referens - File/Reference

    LVR/DTC

    7. Run the command (program)plotta in order to view a figure of the result. One example may be:

    -1 -0.5 0 0.5 110

    20

    30

    40

    50

    60

    70

    80

    90

    log d

    M-C

    8. To display the number of points, simply typeN.

    9. The RMS error between the measured and the calculated values (using the optimised parameters ofA0

    and A1) is given by the formula, ( )RMSN

    L Lcalc predi

    N

    = =

    1 2

    1

    ( Lpred is obtained from SSpred by using

    a simple link budget). To calculate the RMS error, run the command (program) crms.