Long-Distance Running Routes ’ Flat Equivalent Distances from Race
Results and Elevation Profiles
Dimitri de Smet
1
, Michel Verleysen
1
, Marc Francaux
2
and Laure nt Baijot
3
1
ICTEAM, UCLouvain, Louvain-la-Neuve, Belgium
2
IoNS, UCLouvain, Louvain-la-Neuve, Belgium
3
Formyfit, Enghien, Belgium
Keywords:
Equivalent Distance, Endurance, Race Times, Running, Collaborative Filtering, GPS Track, Elevation,
Gradient, Ascent.
Abstract:
Running routes’ elevation profiles affect their runnability and therefore athletes’ average speeds: route dis-
tances alone are not sufficient to predict or evaluate running times. This is an issue for race preparation, race
strategy, performance comparison and runner workload planning anytime the ground is sloping. This paper
proposes a methodology to establish route equivalent distances expressed as a function of their elevation pro-
files. The same expression can be used to compute gradient adjusted speeds either for athlete pacing during
races or to analyze their performances afterward. The approach is first based on race results and addresses the
problem of attendees’ level disparities by evaluating races and athletes at the same time using a race perfor-
mance model. Subsequently, this paper use polynomial and piecewise linear regressions on the instant slope
along the routes to express equivalent distances. They match previous studies with constant slopes and extend
to the case of varying slopes.
1 INTRODUCTION
Race distances alon e are not sufficient to evaluate
athlete race times. For instance, an athlete who runs
a marathon distance in 3:30’ could be a well prepa-
red casual runner if it happened to be at the Berlin’s
Marathon (known to be particu la rly flat) or he could
be a world class champion if it was the Pikes Peak
Marathon (known to be one of world’s toughest ma-
rathon). Elevation gra dient, weather conditions, alti-
tude, vegetation, uneven grou nd and ground firmness
affect a thlete speeds. A flat e quivalent distance, tha t
reflects all race characte ristics is useful in many ways:
athletes can prepare races considering realistic race
lengths. It also makes athlete ranking possible even if
they did not attend to the same races (this also requires
a p erformance model). This paper describes a m e tho-
dology that assigns flat equivalent distances to routes
based, first, on r a ce times and, the n, on their elevation
profiles. The flat equivalent distance is defined as the
distance that would be run , on average, in the same
time if the ground was flat; all other above-mentioned
conditions b eing equal. The methodology is applied
on endurance races ranging from 8 to 159km.
Two paths could be taken to achieve the same
goal: through metabolic m easurements or through
statistics on race results. In the first appro ach, flat
equivalent distance can be computed as the distance
that would lead to the same energy expenditure on flat
ground. The relationship between energy expenditu re
and gradient of ascent is established by (Minetti e t al.,
2002) by measuring athletes oxygen uptake on an in-
clined treadm ill. This first approach is perfectly fine
to assess workload or to plan weight reduction pro-
gram but might not be accurate in what concerns race
times prediction because it does not target it specifi-
cally. The second approach infers a relationship be-
tween race average speeds and their elevation profile
(Kay, 2012; Scarf, 1998; Scarf, 2007). This paper
follows the second approach and addresses two pro-
blems that previous works do n ot take into account.
The first problem is that rac e results ca nnot be ea-
sily c ompared because races are not run by the same
set of athletes. There could be attendee level diffe-
rences that dep e nd on the popularity of the race. For
instance, some local races may fail to attract world
elite runners. On the o ther hand, some very popu lar
races may a ttract a crowd of casual runners. This im-
plies that races results (race recor ds or any race statis-
tics) do n ot necessarily reflect its objective runnabi-
56
Smet, D., Verleysen, M., Francaux, M. and Baijot, L.
Long-Distance Running Routes’ Flat Equivalent Distances from Race Results and Elevation Profiles.
DOI: 10.5220/0006937000560062
In Proceedings of the 6th International Congress on Sport Sciences Research and Technology Support (icSPORTS 2018), pages 56-62
ISBN: 978-989-758-325-4
Copyright © 2018 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
lity. This problem is solved by taking a c ollaborative
filtering app roach, inspired by (de Smet et al., 2017),
that evaluates both athletes and races at the same time.
In this way, equivalent distances approximation can
be computed from a collection o f race results tak ing
the attendees’ level into account.
The second problem is that previous attempts fully
describe race elevation profiles by two global metrics
that are either the cumulative elevation g ain
1
(Scarf,
2007) or the average elevation gradient of non-loop
races (fell running races) that present a relatively co n-
stant gradient (Kay, 2012). This is rarely observed
in practice. In the present p a per, the full elevation
profile is considered; this allows for a more realistic
relationship extraction.
To estab lish the desired relationship between flat
equivalent distances and route elevation profiles, one
need to possess races’ equivalent distances for which
the elevation profiles are kn own. For this purpose,
races equivalent distances are, first, established using
race results. This step is referred to as collaborative
filtering (Section 3) . Then, taking elevation profile
data as inputs, a regression mod el that reproduces the
obtained race equivalent distances is built (Sectio n 4).
These two steps are validated by assessing how equi-
valent distances improve race time prediction compa-
red to actual distances.
In the following, all equations are expressed using
SI unit system: speeds in [m/s], times in [s] and eleva-
tion gr adient in [m/m]. Other u nits are used in figures
for convenience.
2 DATA SOURCES
The two steps methodology requires two kind s of
data. In the first step (the collaborative filterin g part),
race results are used to compute flat equivalent distan-
ces for a set of races. In the second step (the flat equi-
valency modelling part), elevation profiles are used to
model flat equivalent distances as a function of the in-
stant elevation gradient along the routes.
2.1 Race Results
A set of 228 031 races times was gathered b y parsing
official results of 61 6 Belgian races. Th ey represent
a large variety of endurance r aces that took place in
2014 and 2015. From these results a subset of 179
674 race times (445 races, 7480 athletes) is kept to
obtain a data set presenting properties that allow a co l-
laborative filtering a pproach to operate (as explained
1
The cumulative elevation gain is the sum of all positive
vertical displacement along the route.
in Section 3.3). Race results are used to com pute flat
equivalent distances that are then put in relation with
race elevation data.
2.2 Elevation Data
Race routes data were collected through measure-
ments made by runners during the races using their
sports watches. Runners uploaded tracks and made
them publicly available to the online community.
Those tracks contain data such as geographic coor-
dinates, timestamps and altitudes. Consumer grade
GPS-based elevations have poor accuracy (Bauer,
2013). In a previous work (de Sm et et al., 2017) route
elevations we re gathered by querying publicly availa-
ble topography data such a s SRTM data (Shuttle Ra-
dar Topography Mission) or Google Maps APIs. They
are both based on radar topography survey made from
space. It is observed that, in su ch databases, the alti-
tudes of the treetops is assigned to route parts that are
covered by trees: this causes artificial high elevation
gradients on routes that pass under trees.
Fortunately, some high-end sports watches in-
clude barometric altimeter that have good relative
accuracy: the altitude is known with a n additive bias
that would need to be calibrated . The relative accu-
racy is what is of primarily interest, the absolute alti-
tude accuracy being irrelevant to our purpose as our
analysis is based on elevation gradient only. Routes
recorde d with such devices could be found only for
129 races o f our 445. Those 129 races are used to
model our flat equivalent distance model from eleva-
tion data.
Although more tha n two thirds of the races were
not used in the flat equivalency modelling part, they
are still useful in the c ollaborative filtering part be-
cause they help to ch aracterize athletes and there fore
improve the flat equivalent distance estimation of the
129 races that are used in the flat equivalency model-
ling section.
2.3 Instant Elevation Gradient
Elevation profiles as they are record e d, even by high-
end devices, are noisy signals that need to be filtere d;
especially b e cause our application requires to take the
gradient: th e derivative of a noisy signal can take ar-
tificially high amplitudes. A simple way to take the
gradient and smoo thing at the same time is to take the
average altitude on a n-meters distance ahead min us
the average altitude on the same distance behind. The
chosen distance acts then as a smoothing factor. Mo re
formally, if the elevation profile e(x) is re-sampled
every meter, its gradient g(x) at distance x is given
Long-Distance Running Routes’ Flat Equivalent Distances from Race Results and Elevation Profiles
57
Figure 1: Elevation gradient with different smoothing dis-
tances recorded by 6 different sports watches featuring a
barometric alti meter. On the first plot, the elevation gra-
dient is under-smoothed: it shows some details that do not
correlate among measurements. At the opposite, on the last
plot, the elevation gradient is over-smoothed: some details
present in the 6 measurements vanish.
by
g(x) =
i=x+n
i=x+1
e(x)
i=x1
i=xn
e(x)
n
2
, (1)
with n being the smoothing distance in meter. Smoo-
thing reduces the measurement no ise but it also redu-
ces fast changing details of the real gradient. The re-
fore, choosing the smoothing distance n results in a
trade-off that is solved by analyzing its effect on se-
veral track measur ements of the same routes rec orded
by different runner devices. As independence of the
measurement can be assumed, if the filtering distance
is too small, the rando m noise is different on each of
the measuremen t making them less correlated. At the
opposite, a too large smoothing distance makes disap-
pear some details that are present on all the measure-
ment traces. The smoothing problem is illustrated in
Figure 1. Smoothing factor of 120 meters is observed
to be a good trade-off: average tracks pairwise corre-
lation reach a plateau and most small scale elevation
details remain visible.
3 COLLABORATIVE F ILTERIN G
The idea of using c ollaborative filtering presented by
(de Smet et al., 2017) is to take benefit of a collec tion
of race times of many athletes on many races to le-
arn enough information about every race an d every
athlete to best explain race results. The basic under-
lying model is that race times are expla ined by a linear
model of the race paces: the race pace p
a,r
of athlete
a on race r is a sum of products of N race par ameters
r
i
times N athlete parameters a
i
:
p
a,r
=
N
i=1
a
i
· r
i
. (2)
Races and athletes parameters are found by mini-
mizing the quadratic reconstruction error, i. e. the
sum of the square d differences between observed pa-
ces and paces compu te d with Equation (2).
The simplest case is when N = 1 meaning that the
race pac e is equal to a parameter related to the diffi-
culty of the race times a parameter rela ted to the level
of the athlete (low values for high levels). In other
words, g iven all the considered races resu lts, the al-
gorithm outputs race an d athlete parameters that allow
race time predic tion and athlete comparison.
This paper presents a new underlying model that
is more physiologically sou nd and that ma kes the flat
equivalent distance appearing explicitly.
3.1 Race Performances Modelling
Running race times are quite well reflected by a po-
wer law (Garc´ıa-Manso et al., 2012): if elevation va-
riations are small enough to not affect the race time,
an athlete a is expected to run race r w ith an average
speed s that depends on the distance as
log(s) α
1
· log(D
r
) + α
0
(3)
with D
r
being the actual race distance a nd {α
0
,α
1
}
two athlete-specific pa rameters.
Race time t = D
r
/s can be expressed as a function
of a fictive flat equivalent distance D
eq
.
log(t) w
1
· log(D
eq
) + w
0
(4)
with {w
1
,w
0
} = {1 α
1
,α
0
}. Flat equivalent dis-
tances are formally defined a s the distan ces that would
be run in the same time if the ground was flat. They
reflect a ll parameters that may affect the average race
speed.
3.2 Equivalent Distances Estimation
Athlete pa rameters (w
a,0
,w
a,1
) and race equ ivalent
distances D
eq,r
in Equa tion(4) are unknown but race
times must reflect them.
As the number of race re sults is high, athlete par a-
meters and race equivalent d istance can be set so that
the average squared deviation to Equation(4) is mini-
mized on observed race results.
Let log(t
a,r
) be the log race time of athlete a
on race r and log(
˜
t
a,r
) its approximation. Using
Equation(4), they can be expressed as
log(
˜
t
a,r
) = w
a,1
· log(D
eq,r
) + w
a,0
log(t
a,r
). (5)
icSPORTS 2018 - 6th International Congress on Sport Sciences Research and Technology Support
58
The task at hand is now to select race and athlete pa-
rameters such that our predic tion match observed log
race times.
Let be the set of observed (a,r) for which we
have race results t
a,r
. Using the least square error cri-
terion, athlete and race parameters stor e d in vectors
A and R can be selected by solving the optimization
problem
argmin
A,R
(a,r)
(log(t
a,r
) log(
˜
t
a,r
))
2
(6)
This prob le m is solved using an alternate least squa-
res algorithm that alternates between two least squa-
res problems: optimization of the athlete parameters
while holding race parameter constant and optimiza-
tion of the race parameter while holding athle te pa-
rameters constant( Jain et a l., 2013). Convergence is
observed after several iterations, typically 15 to 20.
Solution to Equation (6) does not nece ssarily give
D
eq,r
that correspond to races flat equiva le nt distan-
ces: the optimal solution has one degree of freed om
because athlete parameters w
1
are allowed to compen-
sate freely race parameters D
eq,r
.
Therefore, an extra constraint is required. Some
races present virtually no elevation differences, like
for instance, the Berlin Marathon or some races in co-
astal regions. For races with very flat profile, it can
be expressed that their equivalent distance is equal to
their actual distance:
D
r
= D
eq,r
(7)
For a set
f
of selected flat races, the constraint is
met, on average, if:
f
D
r
=
f
D
eq,r
(8)
3.3 Data Conditioning
The ab ove mentioned constraint assumes that a
a,1
and
D
eq
r
compen sate each other in th e same way for every
pair of (a,r). This is not guara nteed unless races have
several athletes in common. To ensure th at this con -
dition is met. We define a proximity metric between
two races as the number o f a thletes who attend to both
of them. By computing this metric for all pair of ra-
ces, we get an adjacency graph fr om which commu-
nity detec tion algo rithm can select clusters of densely
connected races. Using the Louvain algorithm pre-
sented by (Blondel et al., 2008), only the largest and
most densely connected cluster is kept. The later con-
tain 445 races.
In addition, every athlete must have enough race
results to be characterized (in other words, to be ab le
to set his w
0
and w
1
parameters) even when one or two
race results are kept aside for validation. Therefore,
the collaborative filtering part is applied on athletes
who h ave at le ast 4 race results. This is the case for
7480 of them.
3.4 Validation
The validation of the collaborative filtering part must
assess its ability to a ssign flat equivalent distances to
races. Given the propo sed definition of the flat equi-
valent distance, its quality can be evaluated with the
accuracy of race time prediction.
For this purpose, 1 perc ent of the r ace results are
kept aside for validation. All race results but this 1
percent are used to fit th e pa rameters of each athlete
in Equation (4) with the equivalent distance provided
by the collaborative filtering part. This corresponds to
a simple linear regression per athlete. The accuracy is
eva luated by computing the error at predicting the 1
percent race times tha t were not used while solving
Equation (6). The process is repeated 100 times to
reduce the variance of the error estimate.
4 FLAT EQUIVALENCY
MODELLING
Having flat equivalent distance approximation s for a
set of known ra ces, a regression model can be built to
obtain flat equ ivalent distances based on their eleva-
tion profiles.
Instant elevation gradients for each point on the
race route are computed from the elevation profile as
described in Section 2.3. Let the function F(g) the
distance correction that need to be applied to a given
sloping distance D presenting an elevation gradient g:
D
eq
= D · F(g). (9)
If the considered distance presents a varying gra-
dient, it can be split in 1-meter sub-section x presen-
ting a gradient g(x). The equivalent distance of the
whole route is then the sum of the contributions of
each meter :
D
eq,tot
=
D
x=0
F(g
(x)
), (10)
F(g) can take various forms. The present paper is
restricted to piecewise linear and polynomial functi-
ons. Section 4.1 presents different models that are
found in the literature. Section 4. 2 show how model
coefficients can be fitted to reproduce flat equivalent
distance computed with the techniq ue that is presen-
ted in the collaborative filtering part.
Long-Distance Running Routes’ Flat Equivalent Distances from Race Results and Elevation Profiles
59
4.1 Models
Naismith-like Model
The first rule of thumb, called the Naismith ’s rule, da-
tes from 1892 and is relayed, among others, by (Scarf,
2007). Naismith’s rule formulated in terms of equi-
valent distances in the sense of our definition can be
expressed as
D
eq
D
= F(g) =
(
1, if (g < 0)
1 + f
N
· g, if (g > 0)
(11)
with the f
N
constant being evaluated to 7.92 by (Scarf,
2007).
Polynomial Models
Other papers present a 4
th
or 5
th
order polynomial ex-
pressions that take incre ased spe ed for negative gra-
dients into account. (Minetti et al., 2002) gives a rela-
tion that express the m etabolic energy cost of r unning
by distance unit as a 5
th
order polynomial of the ele-
vation gradient
2
. This equation will be compared to
ours although their definition of eq uivale nt distance in
that c a se would be th e distance that would lead to the
same energy expenditure if it was on flat ground.
(Kay, 2 012) gives a 4
th
order poly nomial that can
be expressed with a definition o f flat equivalent dis-
tance that matches ours.
4.2 Model Fitting
As stated earlier model fittin g assigns model parame-
ter of the unknown function F() of Eq uation (10) so
that it best reproduce flat equivalent distance that are
established solving the o ptimization problem ( 6) dis-
cussed in the collabora tive filtering section.
For the Naismith- like model, fitting Equation (11)
requires to set the p a rameter f
N
. This is done by mini-
mizing the least square error o f the equation by com-
puting, for each races, cumulative elevation g a in.
In previous works, the function F was fitted to
route features using global r oute features e ither by as-
suming a constant gradien t or by taking the cumula-
tive elevation gain. In our case, races present varying
gradient. The general polynomial form can be expres-
sed as
D
eq
D
= F(g) = 1 +
i=P
i=1
f
i
· g
i
(12)
2
metabolic energy cost of running was measured by
quantity of oxygen uptake
with P, the polynomial order. The independent
term is set to 1 because F(g) = 1 for g = 0 : the equi-
valent distance o f a route on flat ground is the distance
itself.
The total equivalent distance (10) can then b e writ-
ten as the sum for each meter x along the route as
D
eq,tot
=
D
x=0
[1 +
i=P
i=1
f
i
· g
i
(x)
], (13)
which can the be re-arranged as
D
eq,tot
= D +
i=P
i=1
[ f
i
D
x=0
g
i
(x)
], (14)
allowing to pre-compute the inner sum for each race
route. The P model parameters f
i
can then be compu-
ted using a simple linear regression .
In the equ ation above the runnability is assumed
to only depend on the elevation g radient. This of
course not completely true in practice. The under-
lying assumption is that all other parameters (like we-
ather c onditions) are independent of the elevation pro-
file so that they will be averaged out in the regression .
4.3 Validation
Just as for the equiva le nt distances provided by
the collaborative filterin g, the quality of the equiva -
lent distance computed using gradient-based formula
F(g) is evaluated with the accuracy of race time pre-
dictions. Again, for eac h athlete, parameters w
a,0
and
w
a,0
in Equation (4) can be set using all race results
for which eq uivale nt distance can b e computed except
1 percent that is kept aside for validation. The error
between validation race results and predicted race re-
sults is computed . The process is also repeated 100
times to reduce the variability of the er ror estimate.
5 RESULTS
This section presents the quality of our flat equivalent
distances as pre dictors fo r athletes’ race times using
the Equation (4). As athletes experien ce high variabi-
lity in their performances (especially casua l runners),
race time predictions can not be highly ac curate: 4 to
6 percents error are observed on average. Neverthe-
less, given our definition of equivalent distance, the
accuracy o f race time prediction is the best way to as-
sess th e quality of the computed equivalent distances.
The boxplot presented in the Figure 2 shows the re-
lative error at predicting athletes’ race times conside-
ring different e quivalent distances. Mode ls expressed
icSPORTS 2018 - 6th International Congress on Sport Sciences Research and Technology Support
60
Figure 2: Tukey boxplot representing relative error while
predicting race times with different equivalent distances.
Computing equivalent distances from race results through
collaborative ltering gives the best r esults.
Figure 3: Distance correction based on the elevation gra-
dient.
by Equation s (11) and (12) are evaluate; both with li-
terature coefficients and with co efficients fitted on our
computed equivalent distances. The Table 1 shows
mean relative errors and model param eters.
The Figure 3 shows Naismith’s piecewise linear
model with original coefficient, two polynomial mo-
dels with original coefficients and our best polyn o-
mial fitting (5
th
order). We can note their relative si-
milarity in th e r a nge that is displayed. As our eleva-
tion profiles do not present much grad ie nt outside of
the range [20%;2 0%], this is also the range of vali-
dity of our model.
6 DISCUSSION
The methodology presented in this paper shows simi-
lar distanc es correction as previous researches with a
novel approa ch that possess some ad vantages.
Table 1: Mean relative errors at predicting race times gi-
ven for different equivalent distances. Actual distances give
7.5%. Collaborative filtering equivalent distances give 4%.
Model Naismith-like 5
th
order polynomial
(Scarf, 2007) (Minetti et al., 2002)
Equation # (11) (12), P = 5
Coefficients f
N
= 7.92 { f
1..5
} =
{5.42,12.9,12.0,
8.44,43.2}
MRE 6.4 % 6.3 %
Model Naismith-like 4
th
order polynomial
(present paper) (Kay, 2012)
Equation # (11) (12), P = 4
Coefficients f
N
= 6.5 { f
1..4
} =
{3.64,17.8,
3.10,23.8}
MRE 6.3 % 6.1 %
Model 5
th
polynomial
(present paper)
Equation # (12), P = 5
Coefficients { f
1..5
} =
{3.90,220,23.6,
6.36,5.34}
MRE 5.8 %
We obtained flat equivalent distances by ap plying
a collaborative filtering technique on race results.
This technique ta kes benefit of p hysiologically sound
power law to evaluate races’ equivalent distance and
athletes’ level at the same time. By doing so, we ad -
dress the problem of athlete level disparity am ong the
different r a ces.
The obtained flat equ ivalent distances ar e used to
build models that take races elevation profile data as
inputs. Unlike previous works, we apply our distance
correction model to routes with varying gradient.
Results prove that the compu te d equivalent d istan-
ces are more relevant than the ac tual distances as race
time predictors. The best distance correction is the
one that is computed on race results because it cap-
tures all parameters that affect race times (elevation
gradient, weather conditions, ground firmness, etc.).
Equivalent distances solely based on the routes’ ele-
vation profiles give all similar improvements.
Races considered in this paper are all loop race s:
they e nd at the same place as they begin. T herefore,
flat eq uiva lency formulas can not be safely generali-
zed to routes for which it is not the case. This is ob-
vious for Naismith-like formulas: in appearance, the
model only considers decreased speed in uphill secti-
ons; but actually, as it depends on race results, the mo-
del coefficient f
N
accounts for the fact that there are
necessarily downhill sections were the speed is at le-
ast a little increased. Indeed, a route with only uphill
sections, would be slower than what is predicted by
the N a ism ith-like fo rmulas that are considered here.
Long-Distance Running Routes’ Flat Equivalent Distances from Race Results and Elevation Profiles
61
The polynomial expression that is obtained for
distance adjustment can serve as-is to correct instant
speed on a race route. This could be used, for in-
stance, to optimiz e r ace m anagemen t by providing a
gradient-adjusted target speed along the route.
REFERENCES
Blondel, V. D., Guillaume, J.-L., Lambiotte, R., and Le-
febvre, E. (2008). Fast unfolding of communities in
large networks. Journal of statistical mechanics: the-
ory and experiment, 2008(10):P10008.
de Smet, D., Verleysen, M., and Francaux, M. (2017). Run-
ning race times prediction and runner performances
comparison using a matrix factorization approach. In
5th International Congress on Sport Sciences Rese-
arch and Technology Support.
Garc´ıa-Manso, J., Mart´ın-Gonz´alez, J., Vaamonde, D., and
Da Silva-Grigoletto, M. (2012). The limitations of
scaling laws in the prediction of performance in endu-
rance events. Journal of theoretical biology, 300:324–
329.
Jain, P., Netrapalli, P., and Sanghavi, S. (2013). Low-rank
matrix completion using alternating minimization. In
Proceedings of the forty-fifth annual ACM symposium
on Theory of computing, pages 665–674. ACM.
Kay, A. (2012). Pace and cri tical gradient for hill runners:
an analysis of race records. Journal of Quantitative
Analysis in Sports, 8(4).
Minetti, A. E., Moia, C. , Roi, G. S., Susta, D., and Ferretti,
G. (2002). Energy cost of walking and running at ex-
treme uphill and downhill slopes. Journal of applied
physiology, 93(3):1039–1046.
Scarf, P. (1998). An empirical basis for naismith’s rule.
Mathematics Today-Bulletin of the I nstitute of Mathe-
matics and its Applications, 34(5):149–152.
Scarf, P. (2007). Route choice in mountain navigation,
naismith’s rule, and the equivalence of distance and
climb. Journal of Sports Sciences, 25(6):719–726.
icSPORTS 2018 - 6th International Congress on Sport Sciences Research and Technology Support
62