The material in this report was addressed to project personnel who already had a basic understanding of the TRACER system. The following should provide context helpful in understing the data in the report. The material should be of interest to those using or contemplating automatic tracking of animals because the accuracy data are valid for any type tracking.

THE SYSTEM

      TRACER was a multiple-tower radio direction finding system intended to keep track of the locations of a Company (225 soldiers) during their land navigation training. TRACER used 6 2-element horizontally polarized Yagi antennas arranged as spokes separated by 60 degrees. A 6-position electronic switch was used to rapidly change antennas. A microprocessor controlled the antennas and a narrow band receiver. Very briefly, the antennas were scanned and signal measured for each antenna. The ratio (R1 in the report) of the strongest to the 2nd strongest signal was used in conjuction with a look-up table to determine the number of degrees the source signal was from the direction the antenna giving the strongest signal was pointed. A second ratio (R2 in the report), that of the strongest to the 3rd strongest, was also obtained. The arrays of 6 Yagis were mounted on towers about 5 meters above the canopy, towers were located at high points in the 10 square km training area (see Fig 1 of the report). The frequency used was 139.450 MHz, similar to the 148 to 168 MHz range used for wildlife tracking.

      TRACER differed from animal tracking in that time-division multiplexing was used to separate the >200 soldier transmitters (all on 139.450 MHz). Wildlife tracking generally uses frequency-division multiplexing to separate individuals (a different frequency for each animal). Another difference was that TRACER could use transmitters of greater power (20 mw radiated) than are usual for animals (typically < 1 mw) because replacing batteries was made routine. Thus, the signal strength for the TRACER tests in the report were considerably stronger than they would be for animal transmitters at similar distances. This signal differential means that the accuracy results in the report would have been worse if the system was tracking animals.

      Data from the various towers was telemetered to a central station where they were displayed and/or filed on an IMB PC-AT.

      The location of towers and of the test transmitter for all data taken was known to the nearest meter thanks to U.S. Army surveyors participation in placement of the many "staked" locations used for the tests.

PERFORMANCE

      I will briefly describe calibration attempts that preceded the various test results in the report. It being anticipated that the terrain and vegetation would make accurate system calibration impossible from test transmitters near the ground, a remotely operable test transmitter was installed near the top of each tower as part of the system. The signal path between all of the tower tops was line-of-sight. At that point in the project I viewed "calibration" a matter of aligning the antenna arrays physically so that the north antenna pointed exactly north, etc. The arrays had already been calibrated to a look-up table for signal ratios R1 and R2 as mentioned above. The bearing data from each of the 7 towers to the others (42 bearings in all) had a lot of scatter, as if the towers were not where they were surveyed; actually worse than that, as if the towers were in more than one location. Consistency was limited to about the nearest 5 degrees. We concluded that even the line-of-sight propagation was distorted by the underlying terrain and vegetation. Still hoping for a definitive calibration, we put a test transmitter on a balloon 30 to 50 meters above the canopy at a known location over high ground. After "calibrating" the towers to this, the balloon was moved to a new location and bearings taken. Again, there was no consistency. The measured and expected bearings were different, not all of them at any one location, but always some of them. Even during calibration at the first location, we knew there would be a problem because the balloon varied its position in the wind by about 5 meters, a distance that subtended about 1 degree at the closest towers. Yet the measured bearings were varying by about 3 degrees, more for some towers, less for others. We ended up using the average of the 6 bearings taken to the "other" 6 towers for physically aligning an array on a tower. Further testing (including data presented in the report) revealed that biases for a given calibration varied with azimuth, ie., accuracy could be gained by using an azimuth dependent correction factor. These findings alone showed that published methods of calibration for systems using fixed location DF antennas are little better than no calibration at all (see White ............).

      No details are given in the report regarding the analyses program. Two plotting algorithms were used. "SIMPLE" refers to using the centroid of the polygon formed by the intersection of bearings from the multiple towers. "WEIGHTED" refers to a more complex algorithm as follows. The bearing from each tower was evaluated in terms of the information it provided for north-south positioning and for east-west positioning. For instance, a north (000) bearing provides no N-S information and maximum E-W information. Sine (for N-S) and cosine (for E-W) correctly weight bearings in this way. The simple location, as described above, gave a first estimate of location from which a first estimate of distance from each tower was calculated. The combined E-W (and N-S) weighting for a bearing was its cosine (and sine) divided by the first distance estimate. The latter necessary because a given angle subtends an arc length proportional to distance. We found, however, that at distances less than 200 m the increased weighting was not valid because of scattering of signal in the vicinity of the tower. We therefore truncated the distance function at 200 m. Points were then triangulated using all the possible pairs of bearings, with each point given E-W and N-S weights proportional to the product of the E-W and N-S weights of the 2 bearings from which the point was triangulated. With multiple towers, many such points were calculated for each fix, each providing a weighted E-W and N-S coordinate. The weighted E-W coordinates for all points were averaged to determine a final E-W coordinate, likewise N-S for a final N-S coordinate. The final location estimate was given by these coordinates. The number of points depended on the number of towers giving bearings: 2 towers give 1 point, 3 give 3, 4 give 6, 5 give 10, 6 give 15, 7 give 21, etc. A second iteration (based upon a refined distance estimate and recalculation as above) was found not to improve accuracy and was therefore not used.

      For both algorithms, points were thrown out if they did not meet two criteria (see the report under PLOTTING for the criteria CRIT and ANG). The programs included these criteria as variables and the values finally used in the report were based on multiple runs that revealed the variations of yield (ratio of used to total fixes) to reduction in mean fix error.

      Yield was also a major consideration in what the report calls "outlier-kill" which worked as follows. A final fix was determined as above (either simple or weighted). The distance was calculated to the fix point from each of the points from which the fix was derived. The point with the greatest distance was then thrown out and the calculation of a fix repeated but without that point. This could be repeated to remove the 2nd worse point, the 3rd, etc., but with ANG and CRIT already reducing points by throwing out tower bearings, the yield would have been reduced to useless for this application by loss of more points with outlier kill. Many other schemes were tried, some working with points, some directly with bearings, some with coordinates separately, but none worked better, or for that matter hardly worse, than the above, ie, none gave much improvement (see Table 2 in the report).

      I recall the frustration, now 15-years old, in knowing the location, plotting the point, and so often seeing just 1 or 2 of 6 or 7 bearings reduce accuracy markedly. Thus began the search for a way to indentify these bad bearings. Armed with the large data set, I was at first confident, but after months of work the best I could do is given in the report. The potential improvement, even if the so-called bad bearings could be identified, was limited to something around a factor of 2 (see Fig 10 in the report).

      All the above apply to single fixes from single bearings from multiple towers. We early noted that large changes in bearings often resulted from movements of the transmitter as small as 1 meter. Much of the report concerns improvements made by averaging bearings taken in succession to a slowly moving transmitter. This approach proved to be of academic interest because the soldiers would often stop to study a map or take a compass bearing or just rest or sometimes move so fast that movement error would overide fix error. Similar considerations apply to animals. However, in the frequency-division multiplexing scheme employed with animals, a system has complete control of the times fixes are taken.

      The phenomenon of small movement causing large bearing differences is due to multipath propagation changes with movements of the order of a half wavelength or more at the frequency in use (1.1 meter at 139 MHz). The multipath environment would be different for horizontal and vertical polarization which led to the suggestion in the report of using both simultaneously, thereby giving the equivalent of moving the transmitter once to provide 2 bearings to average. But this would at best have provided about 25% improvement in accuracy (Table 2 in the report), hardly enough to meet the 15-meter accuracy required of the system.

      TRACER thus died from vegetation and terrain induced multipath problems. Everything else worked, including taking 8 bearings per second from up to 240 transmitters on the same frequency and streaming these data from 7 towers to an IBM PC-AT that filed the data and plotted the fixes. Subsequently (1990-93 and 1996-97) I used a similar system to track animals. In 1992 the system was independently tested and described by Larkin, Diehl, and Raim (1996. Journal of Field Ornithology. volume 67(1) pp 59-71). Although I appreciate their enthusiastic attitude, my years of experience, including TRACER, suggest that they were somewhat overly optimistic about what such a system can do (see Cochran, Swenson and Pater, Radio Direction-finding for Wildlife Research).

Figure 0. A TRACER plot of a walk along Yankee control road. The center of the small squares (30 by 30 meters) are points where the transmitter was triggered; dots on the line are corresponding TRACER-plotted locations. The mean error for the points is 40 meters. The border is 2 km (scale = 1:25,000) with the lower left corner at grid coordinates 104,000 east and 77,000 north.

INTRODUCTION

      During November and early December, 1987, most the TRACER hardware was operational and field tests were conducted. During December and January, I made various analyses of the data files generated by these and other tests and found evidence that plotting errors are large, random, and terrain-induced. In fact, so large and so random that only the crudest equipment calibration was needed. Whereas, in the conceptual stage we were dedicated to refining the equipment to provide fractional degree accuracy, in operation we are repeatedly confronted with 3, 5, 10, and sometimes 20-degree terrain-induced errors.

      Very briefly, the mean accuracy of TRACER is about 69 meters. By using a moving average method, the mean accuracy can be reduced to about 45 meters. These conclusions are based on field data from 26 locations along or near Yankee control road. The primary impediment to accuracy is random scattering of the signal in the general area of the transmitter. A bearing error distribution taken in flat, treeless Central Illinois terrain indicates that a TRACER installation there would provide plotting errors one fifth to one tenth of those given above.

THE TESTS

      Grids at known locations were established (Fig. 1). With Yankee stakes A through Q as centers, a 20 by 20-meter grid was made around the centers. Each grid consists of 9 points: 1 point at the center, 4 points 10 meters N, E, S, & W of center, and 4 points 14.1 meters NW, NE, SE, & SW of center. Two additional pseudo stakes at points between stakes D & E and between I & J were gridded in the same way for a total of 19 locations with 20 by 20 grids.

Fig. 1. Map of the Yankee area showing approximate locations of places along Yankee control road from which transmissions were made. The small black squares (20 by 10 meter grids of 9 points) are centered on stakes A through Q and on two in-between places: ED and IJ. ED is incorrectly positioned 60 meters west. The large square is a 100 by 100 meter grid of 100 points. Towers are located at crosses.

      One 100 by 100-meter N-S, E-W oriented grid was established with stake K at its southeast corner (Figs. 1 & 2). This consisted of 100 points with 10-meter spacing. A supplementary 100 by 100 grid used points 2.5 meters north of the points in the above grid.

      The test transmitter, with its antenna at waist height, was triggered several times at each of the points in all the grids. In addition, at the 100 X 100 supplementary grid, the test transmitter was also triggered with its antenna at head-height.

      Tower 0 (Crosbie) was not used in any of the tests. When the first 100 by 100 grid was used, tower 7 (Yankee Bl.) was not functioning; when the supplementary 100 by 100 grid was used, the gray box from tower 5 had been moved to tower 7 so that tower 5 was not operating. At stakes A to Q and at DE and IJ, all 7 towers were operating.

THE DATA SETS

Fig. 3 Scatter plots of actual bearing (horizontal axis) and measured bearing (vertical axis). Tick marks are at 5-degree intervals. Diagonal line is where all points would lie if all bearings were perfect. The bearings were to various points along Yankee control road from stake A to stake Q.

Fig. 4. Same as Fig. 3 except different towers. Tower 8 is not literally a tower site; it was the system set up at Blue Field.

      The scatter plots (Figs. 3 & 4) are from the 20 by 20 grids A through Q. The data were sorted so that no more than one set was included from each point. Frequency distributions of errors (Figs. 5 & 6) and course plots (Figs. 7 & 8) are from 20 by 20 grid data as above, but with an added nine 20 by 20 grids using data from the 100 by 100 grids. Analysis of bearing variations due to movement (Table 1) used all points from which at least 5 transmissions were made.

Fig. 5 Frequency distributions of error for the seven towers separately (a-g) and all towers combined and normalized for zero error (h). Vertical axis is percent of sample (marks at 2.5% for a through g, and 5% for h. Horizontal axis is error in degrees; marks at 5 degrees, samples at (+ -) 0 to 0.5, 0.5 to 1.0, etc.

Fig. 6 Frequency distributions of error with cumulative increase in nearest-neighbor averaging of bearings. Bearings were taken in Yankee area at Ft. Benning (a-f) and in open flat land in Illinois (g). Upper left (a) is same as Fig. 5 lower right (h).

Fig. 6 f and e.

Fig. 7 Plots from 100 by 100 meter grid data. Border is 2 km with lower left corner at E-W 104000 and N-S 77000 grid coordinates; scale is 1:25,000. The actual sequence of transmitter location was at the center of the boxes starting from the SW corner, then to NW to NE to Se to S to middle. The left column of plots was made using one bearing from each tower. They differ in that bearings to different points in the 20 by 20 sub grids (of the 100 by 100) were used. The program centers the 30 by 30 boxes on the point, or mean of the points, actually used. The upper right plot was made using the average of the six bearings were averaged. For the six plots to the left the mean error was 77 meters; for top right mean error was 39 meters, for bottom right 30 meters.

Fig. 8 Plots along Yankee control road and around the outside of the 100 by 100 grid. The scale is the same as in Fig. 7. At bottom right one bearing from each tower was used, i.e., no averaging. Top left plotted using 3-bearing average, bottom left uses 6-bearing average, and top right uses a 9-bearing average. In all cases, points were inside the 30 by 30 meter boxes that represent the actual route.

BEARINGS - THE RAW DATA FROM TRACER

      Scatter plots of measured versus actual bearings to the points in al 20 by 20 grids are shown in figures 3 and 4. The actual bearings are computed n the basis of stake coordinate lists and measurements of the tower coordinates by PAL (see Appendix 1). If all bearings were perfect, except for a constant due to mechanical mis-positioning of the antenna arrays, all points would plot on a line parallel to the diagonal line, i.e., measured bearings would always equal actual (true) bearings plus some tower specific constant. That is certainly not the case. Figure 4 shows data for tower 8 which was the system set up in the open on Blue Field. Although the sample is small, variability is much less than for data taken in the vegetated and hilly terrain in the Yankee area.

      Another way of examining the data of figures 3 and 4 is by frequency distributions of bearing errors (Figs 5 & 6). In the open, the distribution of bearing errors is tightly grouped around zero (Fig. 6g: errors of 1 degree or less comprised 92.5% of the sample). In contrast, in the Yankee terrain, only 32.6% of the sample was 1 degree or less in error (Fig. 5h, 6a). Individual distributions for each tower (Fig. 5a-g) differ from that of the lumped data (Fig. 5h) in a manner that reflects their smaller sample size and varying distance to the transmitter.

      Distance between tower and transmitter affected the error distributions in a manner consistent with the hypothesis that most error is generated by scattering in the vicinity of the transmitter. This scattering error is an order of magnitude or more greater than that contributed by inaccuracies in the bearing measurement. Distances from tower 7 to the transmitter locations were least; those from tower 5 greatest. Correspondingly, the spread in the error distribution for tower 7 was about twice that of tower 5 (Fig. 5e, 5g). However, tower 1 (Fig. 5a) was almost as far from the transmitter as tower 7, yet, the error for tower 1 is noticeably greater. This suggests another non-equipment error source, namely the intervening terrain. Tower 1 looks over a ridge to Yankee control road, tower 7 does not (Fig. 1).

AVERAGING BEARINGS TO IMPROVE ACCURACY

      The lumped sample (Figs. 5h & 6a) includes data from all the towers and from points at many locations. Such a segmented sample is useful for examining variation. By averaging nearest neighbor bearings (with an average calibration constant applied to each tower to normalize their summing to zero error), the error distributions become tighter (fig. 6b to 6f). For example, by averaging 20 bearings (Fig. 6f), 69.1% of the averaged bearings are in error by 1 degree or less. This is a considerable improvement over 32.6% without averaging (Figs. 5h & 6a), but is still much worse than the 92.5% (Fig. 6h) for operation in open, flat land.

      How far apart must points be in order that the average bearing to their average location is better than any of the individual bearings to their individual locations? The data sets include four to seven bearings from each of the points in the various grids; also, the 100 by 100 supplementary grid included data for waist and head height at the same points (about 1 meter vertical separation). From these data we find variation among numerous measured bearings to the same point to be quite small (<1 degree for 90% of the sample, Table 1) and dramatically larger for a movement as small as 1 meter. Examination of Table 1 suggests that point separations as little as 1 meter may be adequate for effective averaging and that for optimum averaging separations something less than 10 meters would suffice. The gain in accuracy obtained from averaged bearings is illustrated in the improved plotting accuracy thus obtained (see below).

PLOTTING

      Although bearings are the key to understanding limitations and possibilities, performance of a system like TRACER is better illustrated by how well a route can be recreated on a map. Table 2 summarizes the mean metric error in each of the graphic plots (figs. 7 & 8). Simple plotting uses the simple mean of all points generated by crossing of bearing lines of all towers, subject to ANG limits (see below). WEIGHTED plotting takes into account distance and the bearings' contribution to E-W and N-S positioning.

      In all plotting with more than one bearing from each tower averaged, a maximum range allowance (MAXR) of 24 degrees was applied. In all plotting, the following apply. Data were not used if bearings computed from R1 and R2 differed by more than (CRIT) 6 degrees. Minimum allowed crossing angle (ANG) was also specified; its optimum value was found to depend on the number of bearings averaged; 10 degrees was a good all purpose value.

IMPROVING ACCURACY

      An effective way to improve accuracy, as illustrated above, is to increase the sample of bearings taken at points several or more meters apart. The key to averaging is how often a transmitter comes on and how fast the subject walks. With TRACER set for an 8-second cycle time, averaging could be done on 3 or 4 and possibly 8 points spaced about 10 meters (the distance walked in 8 seconds). Results should be similar to those of figure 8 (upper or lower left). Unfortunately, the number of individuals that can be simultaneously tracked with 8-second cycle time is only about 30. The next currently available cycle time is 24 seconds which would limit averaging to 2 points.

      In addition to averaging bearings, removal of outlier points provides improvement (Table 2 and plot on cover page). Another way to increase accuracy is to use more towers (Table 4). Unfortunately, bearing sample size and number of towers have a complex and interactive affect on the effectiveness of outlier-removal and plotting schemes, therefore, analyses over the full range of the many variables has not been attempted.

CONCLUSION

        I have done what I could with a data set consisting of a couple dozen box grids along a ridge. The tables, plots, and tentative conclusions are thus limited. We need a few more data sets consisting of linear walks on routes in varying terrain and one or two sets form grids with 2- or 3- meter spacing. Having said that, I conclude that the accuracy of TRACER is now limited to a mean error of 65 to 70 meters for plotting without averaging, and that mean accuracy can be improved to between 45 and 30 meters using a combination of averaging and perhaps more towers.

RECOMMENDATIONS

        The difficulty for TRACER is illustrated in figure 9, i.e., what trees and hills do to the path of VHF radio waves. The difference is even worse than it appears because a number of good bearings are required, not just one. The areas under the curves can be considered 1 and the probability of an error of N (an integer) degrees is the vertical distance to the curve. The vertical ticks are then .05. For the upper curve, the probability of an error less than 1.5 degrees (-1.5 to +1.5) is about .52 + .21 +.19 = 0.92. The probability of getting four such bearings at one swoop is 0.92 to the fourth power or 0.71; almost three chances out of four are not bad. In contrast, for the lower curve, in the hills and trees we get .12 + .11 + .10 = 0.33. To get four such bearings, 0.33 to the fourth is 0.01; one chance in a hundred. But that did, more or less, happen: see the lower left plot of figure 7, the middle left box has a point in it!

        The classic maneuver if you are stuck with "noisy" data (and we are) is to collect a lot of it (get the big sample) and then apply statistics to get a satisfactory estimate of the underlying truth. That is, of course, what the analyses above have shown to be effective. On average, during the tests we had 6.5 towers operational. In practice there should be 8 towers operational. That will help. The outly-kill program was written after the plots were made and I chose not to redo them. I have, however, made one plot using this program with a moving average of 8 bearings. I believe 8 is a practical number for soldiers if they don't run.

        This plot is on the cover. Assume that we can do this well but no better and decide whether or not such plots will be useful. If such plots would be useful, and if work on TRACER can continue, I recommend (1) bringing the system to full operation as is, (2) operating it with real people, and (3) if performance matches the assumption above, continuing background research into ways to improve its performance. I will apply myself to outlining a program for 1 to 3 above when the double "if" is resolved.

Fig. 9 Same as Fig. 6a & 6g; to compare the scattering effect of vegetation and/or terrain (lower plot) with operation in flat open areas (upper plot). Vertical scale is probability of bearing error for unit degree intervals (horizontal scale).

SUPPLEMENT, 12 Feb, 1988

FILTERING OUT THE WORST BEARINGS

        Since ratio 1 (R1) and ratio 2 (R2) are independent measures of bearing, a way to flag bad bearings is to compare the bearings derived from R1 and R2 and throw both out if they differ by too much. In the preceding discussion CRIT was the variable that described "too much". For CRIT values larger than 8 degrees, the ratio of good bearings discarded to bad bearings discarded approaches the ratio of good-to-bad bearings left in the sample. Therefore, larger values of CRIT reduce the sample size without a net benefit, and since sample size is a factor in accuracy, accuracy will decrease. Keep in mind that setting CRIT to 8 degrees in the programs does not mean that bearings with errors greater than 8 degrees will be removed from the sample; indeed, errors of 30 and 40 degrees are in the samples analyzed in spite of CRIT. In fact, these large errors are responsible for most of the very poor fixes.

        Signal strength was found not to be related to the accuracy of a bearing if the signal was stronger than about 16 (about 0.3 microvolts). All analyses presented were for data with signals less than 16 filtered out.

        With the existing hardware and firmware, the above described filtering is alL that is available. It provides some improvement (10 to 20% of the really large errors are detected). A significant improvement would result if there were a way to detect a substantial percentage of the many remaining large errors.

BENEFITS OF IMPROVED FILTERING

        Before suggesting firmware or hardware changes that might improve flagging of large bearing errors, I analyzed the effect of such removal (Fig. 10). Nine sample sets were run through the plotting program modified to discard (without replacement) bearings in error greater than a constant (the constant was varied - horizontal axis). The mean metric error and range and the mean maximum error and its range decrease (vertical axis) as bearing error is limited to smaller values. In effect, the program is chopping off the skirts of the frequency distribution of bearing error curves (see Fig. 5h). Theoretically, the curves should extrapolate through the origin (zero error, dotted line), but the discarded data are not replaced so the sample of bearings (number of towers used for a fix) decreases, causing an increase in error (see Table 4) that partially offsets the improvement due to the improved quality of bearings. With a 3-degree cutoff, the mean number of towers used has dropped from about 6.5 to 3; by itself, this reduction would degrade mean accuracy to about 90 meters (Table 4), but the quality of the remaining data is so much better that an increase in accuracy to about 40 meters is achieved. If we had say 10 towers to start with, a 3-degree cutoff should reduce mean error to perhaps 25 meters (approaching the dotted line in figure 10).

Figure 10. Vertical axis is the mean plotting error (lower curve) and mean maximum plotting error (upper curve) resulting after bearing errors greater than selected values (horizontal axis) cause removal of the bearing, without replacement, form the set from which a fix is determined. Data are for 9 fixes each for 26 locations (Figs. 1 and 2) using, at most, 1 bearing per tower (per fix) if the error of the bearing is less than the indicated value (horizontal axis); if greater than the indicated value, no bearing was used for that particular tower.

WILL VERTICAL POLARIZATIN MEASUREMENTS HELP?

        In a 1986 test, it was found that horizontal polarization was, on average, superior (by a factor of about 1.6) to vertical polarization in yielding good bearings. Therefore, we configured the system for horizontal polarization.

        Regardless of how the transmitter antenna is configured, a substantial mixing of polarization occurs in the vicinity of the transmitter due to chaotic reflection and diffraction from vegetation and terrain. Thus, two independent wavefronts travel toward tower sites and two independent bearings could be obtained if tower sites were equipped to DF on vertically polarized signals. Although we have already determined that he horizontal wavefront is more reliable, there is a possibility that the vertical measurement would serve as an effective check for large errors.

HOW TO FIND OUT

        Since system performance can be roughly predicted from a frequency distribution of bearing error (see Figs. 5 and 6), we need equip only one tower site for vertical polarization and then obtain a large sample of bearings to test filtering schemes using the added vertical-polarization data.

        I suggest that one tower site be selected (say Yankee bleacher because it has easy access) and the equipment from Crosbie added at that site temporarily. The Crosbie antenna can be configured for vertical polarization by loosening the clamps and rotating the antennas 90 degrees. This antenna would be mounted above existing horizontal antennas. Crosbie cables are short so it would be necessary to set the gray box on a shelf part way up Yankee tower. The Crosbie data-link transmitter and antenna would also be needed. In effect, we operate two entirely independent units at one tower. Prior to setting up the Crosbie equipment at Yankee, it should be operated (on the short self-standing tower in the BOO room) in the middle of Blue Field for calibration (the calibration for vertically mounted antennas will be different).

        I know that some improvement will result. If it turns out to be equivalent to going from Fig. 6a to 6b., I would hesitate to call it significant. Going from Fig. 6a to 6c would definitely be significant.

IF IT WORKS, WHAT THEN?

        If the use of both polarizations provides a useful improvement, a permanent conversion would not be too difficult. Of course, we would not simply double up equipment as I have suggested for the test. Instead, we would add a vertical set of elements on the booms (radial pipes) now supporting the existing horizontal antennas and add another antenna switch to control the vertical antennas. One gray box (with modified firmware) would control both antenna arrays. No additional cabling would be needed since control and feed from the vertical antennas would be through the existing cables. IN the universe of possible modifications to TRACER, this one is the neatest and simplest, requiring no new designs and a modest change in firmware. Only the mechanical means of mounting the vertical elements to the existing booms needs to be thought out. In the universe of possible modifications to TRACER, this one is the neatest and simplest, requiring no new designs and a modest change in firmware. Only the mechanical means of mounting the vertical elements to the existing booms needs to be thought out. Work yes, research and development no!

CONCLUSION

           Keep in mind that Figure 10 is from data without the benefit of any averaging. Use of a moving average of only 2 bearings improves mean accuracy from 69 to 51 meters (Table 2). Figure 10 indicates that if we can detect errors greater than about 6 degrees, the mean accuracy would improve from 69 to about 50 meters. I don't believe that improvements from these two tactics would add linearly, but they would add and I would hope to see an overall mean accuracy of say 40 meters result. To me, this appears to be a realistic and worthwhile goal to strive for.