Planning Training Loads to Develop Technique and Rhythm in the 400 m

Hurdles using RBF Network

Krzysztof Przednowek

, Janusz Iskra

, Stanislaw Cieszkowski

and Karolina H. Przednowek

Faculty of Physical Education, University of Rzeszow, Rzeszow, Poland

Faculty of Physical Education and Physiotherapy, Opole University of Technology, Opole, Poland

Keywords:

Hurdle Races, Technique In Hurdles, Artiﬁcial Neural Networks.

Abstract:

In this paper training loads to develop technique and rhythm in hurdles are presented. The training loads were

generated using an artiﬁcial neural networks model with radial basis functions. The analysis included 21 hur-

dlers who were members of the Polish National Team. The calculations for the neural model were made using

48 training programmes. The evaluation of the models was carried out using the cross-validation method.

Five independent variables (age, body height, body weight, current result and expected result) and four de-

pendent variables representing the selected training loads were analyzed. The determined model generated

training loads with an error of approximately 21%. Experimental results showed the training programme for

a hypothetical athlete. The analysis shows that all the examined training loads are of a non-linear nature. The

proposed solution can be used as a tool to support planning for selected training loads in 400 m hurdles.

1 INTRODUCTION

Hurdles races are complex athletic events since they

require both motor and technical skills. The re-

sults achieved in these races depend on the level of

strength, the jumping technique and the so-called hur-

dle rhythm (McFarlane, 2000). In the 400 m hurdles

technique plays an exceptional role. The use of ap-

propriate technical skills while taking off, jumping

over the hurdle and landing, is very important. The

400 m hurdles technique is usually referred to as that

of jumping 10 hurdles, each 91.4 cm high. It includes

the individual stages of the race, i.e. start, racing to

the ﬁrst hurdle, racing through the hurdles and rac-

ing to the ﬁnish line. Hurdling, and strictly speaking

jumping over the hurdle, is a form of complex, dy-

namic motion, described in studies as a classic exam-

ple of using the laws of physics in sport (Iskra, 2012).

The evaluation of race technique comes down

to the biomechanical assessment of each individual

component (

Coh et al., 2008). In the course of biome-

chanical analysis, errors in movement are discovered

and can be subsequently corrected by means of an ap-

propriate training plan. While planning the training

loads, the coach very often relies exclusively on his

own expertise. Such an approach sometimes lacks

scientiﬁc basis. It is therefore necessary to look for

solutions that would support the planning of train-

ing loads. One such solution may be the application

of advanced mathematical models (Maszczyk et al.,

2014; Wiktorowicz et al., 2015). Using these tech-

niques leads to a better understanding of the subject

under consideration. The most commonly used meth-

ods of mathematical support for the process of sports

training include artiﬁcial neural networks (Ryguła,

2005; Pfeiffer and Perl, 2006; Perl et al., 2013; Silva

et al., 2007). In sports science neural models are

widely used for modelling, prediction and optimiza-

tion. These models make it possible to predict sport-

ing talent (Roczniok et al., 2007) or determine the

impact of the training on the result achieved (Przed-

nowek et al., 2014). In this study we therefore, de-

cided to use artiﬁcial neural networks in planning the

training loads to develop technique and rhythm.

A novel approach to planning training loads de-

veloped by the authors, is the construction of model-

generated training loads using selected parameters

characterizing the athlete and his current results. This

supports the planned training programme in the train-

ing period under consideration (special preparation

period). The main purpose of this study is the con-

struction of artiﬁcial neural networks to generate the

training loads to develop selected technique compo-

nents in a 400 m hurdles race. The construction of the

model was based on training data from athletes with

a high level of ﬁtness.

Przednowek, K., Iskra, J., Cieszkowski, S. and Przednowek, K..

Planning Training Loads to Develop Technique and Rhythm in the 400m Hurdles using RBF Network.

In Proceedings of the 3rd International Congress on Sport Sciences Research and Technology Support (icSPORTS 2015), pages 245-249

ISBN: 978-989-758-159-5

245

2 MATERIAL AND METHODS

2.1 Training Data

The training data used for the construction of the

training planning model were taken from athletes

competing at a high level. The analysis included

21 Polish hurdlers who were members of the Polish

National Team and represented Poland at Olympic

Games, and European and World Championships.

The 48 training programmes carried out during the

special preparation period were selected. Addition-

ally the results before and after the analyzed training

period were registered. The special preparation pe-

riod usually lasts about three months (from February

to May). Five independent variables (x

– age, x

–

body height, x

– body weight, x

– current result, x

– expected result) and four dependent variables (y

, y

) representing the training loads were ana-

lyzed. A training load is the work or exercise that an

athlete performs during a training session. The se-

lected training loads are those loads which make up

technique and rhythm (Tab. 1). The values of these

training loads are measured in a number of races. The

basic statistic ( ¯x – mean value, min – minimum, max

– maximum, sd - standard deviation) of the variables

used to calculate the model are presented in Table 1.

Due to the difﬁculties connected with carrying

out the test in 400 m hurdles races during the special

preparation period (winter), the athletes ran a 500 m

test race (ﬂat run). As demonstrated in the previous

study, the result of a 500 m race reﬂect the hurdler’s

current performance in a 400 m race (Alejo, 1993;

Przednowek et al., 2014). In this study, the 500 m ﬂat

run was adopted as an indicator of ﬁtness level.

2.2 The Idea of Supporting the Training

Process

The proposed method to support the training process

involves the use of a mathematical model to gener-

ate training loads with given input parameters (inde-

pendent variables). The coach using the model inputs

age, body weight and body height statistics for the

competitor (Fig. 1). At this stage his current result

for the 500 m race is entered, which reﬂects his cur-

rent physical condition and the result expected. At the

output stage of the model the values of the training

loads are generated (Fig. 1). The loads thus gener-

ated constitute a training plan to prepare the hurdler

for the special preparation period. The values appear-

ing at outputs y

–y

represent the sum of all loads of

that type, which should be implemented during the

entire training period. Based on the suggestions from

Generating

of training

loads

Figure 1: Block diagram of models generating training

loads.

the system, the coach plans the training loads to be

carried out each day during the special preparation pe-

riod.

2.3 Calculating and Evaluating Method

In the conducted analysis, the model of an artiﬁcial

neural network with radial basis functions (RBF) was

applied (Bishop, 2006). Networks with a radial ba-

sis functions have one hidden layer, composed of ra-

dial neurons and an output layer consisting of linear

neurons. The RBF networks were implemented us-

ing the Statistica 10 software (StatSoft, Inc., 2011).

In the process of ﬁnding the best model, networks

with various numbers of neurons in the hidden layer

(from 0 to 10) were analyzed. During the evalua-

tion of the neural network, the leave-one-out cross-

validation method was used (James et al., 2013);

cross-validation error was deﬁned as:

∑

i=1

i j

− ˆy

−i j

)

max(y

) − min(y

)

· 100, (1)

where: n – number of patterns (48), y

i j

– real value,

ˆy

−i j

– the output value constructed in i–th step of

cross-validation based on a data set containing no test-

ing pair (x

), CV

– cross validation error for j–th

output. The main criterion for model selection was

the arithmetic error average, calculated for all net-

work outputs. The cross-validation was implemented

using Visual Basic language.

3 RESULTS

The research results are demonstrated in two sections.

In the ﬁrst one, the model calculation is presented,

while in the second section the generated training

loads are analyzed.

icSPORTS 2015 - International Congress on Sport Sciences Research and Technology Support

246

Table 1: Description of the variables.

Variable Description ¯x min max sd

Expected results on 500 m run (s) 65.06 61.50 69.10 1.80

Age (years) 22.25 19.00 27.00 1.97

Body height (cm) 185.04 177.00 192.00 4.70

Body weight (kg) 74.29 69.00 82.00 2.71

Current results on 500 m run (s) 66.78 62.50 71.15 1.68

Runs over 1–3 hurdles (amount) 46.40 3.00 148.00 28.87

Runs over 4–7 hurdles(amount) 82.38 4.00 176.00 40.23

Runs over 8–12 hurdles (amount) 79.71 0.00 194.00 45.83

Hurdle runs in varied rhythm (amount) 330.02 0.00 745.00 156.35

2 4 6 8 10

21.5 22.0 22.5 23.0 23.5

Number of neurons in hidden layer

Figure 2: Mean cross-validation errors; The X-axis repre-

sents the number of neurons in the hidden layer of RBF

network which range from 1 to 10. The Y-axis represents

the mean value of CV

error for all outputs.

0 5 10 15 20

20.14

22.61

21.58

21.16

Figure 3: Cross-validation errors CV

of each output.

3.1 Model Calculation

In order to determine the network featuring the best

generalization ability, a cross-validation was per-

formed. Networks with hidden neurons from 1 to 10

were examined. The cross-validation results are pre-

sented in Figure 2.

The conducted analysis shows that the best model

is the artiﬁcial neural network with seven neurons in

the hidden layer. That network generates an average

cross-validation error of 21%. Errors generated by in-

dividual network outputs are presented in Figure 3.

Output y

is characterized by the smallest generaliza-

tion error (20.14%), while the y

output features the

largest error (22.61%).

3.2 Generation of Training Loads

The next step in the analysis was to calculate the train-

ing loads using the selected RBF network. On the

network input, the data from a hypothetical athlete

(age 21, body height 185 cm, weight 75 kg) were en-

tered. Loads were generated on the assumption that

the 500 m race results would be improved by one sec-

ond, taking as the output result, results from 68 s to

62 s, respectively. The range of result from 68 s to

62 s reﬂects the career of a hurdler. The results of

this experiment are presented in Figure 4. The graphs

show loads generated in such a way that the Y-axis

represents the level of training loads while the X-axis

is the expected result. The values on the X-axis are

placed in descending order as the increase in the com-

petitors’ sports level is associated with a decrease in

the time they achieve over a speciﬁed distance. For

example, if a competitor wants to improve his result

from 66 s to 65 s then the generated training plan indi-

cates that the competitor should implement a workout

with the following capacity: y

= 54; y

= 91; y

= 87

and y

= 225.

The analysis of training loads generated for hypo-

thetical athletes shows that all the examined loads are

of non-linear nature (Fig. 4). Considering the values

calculated for y

(Fig. 4(a)), it should be noted that as

well as the achievement of better results, the volume

of the loads increases. The maximum value of y

observed when the athlete demonstrates a high level

of ﬁtness.

A different trend is observed for the values gener-

ated at the y

output (Fig. 4(b)). Initially, there is a

slight increase of the value and stabilization at sports

level of 65–64 seconds. In the later stages of a ca-

reer it can be seen that as the athlete’s sports level

increases, the y

value decreases.

Planning Training Loads to Develop Technique and Rhythm in the 400m Hurdles using RBF Network

247

66 65 64 63 62 61

50 60 70 80 90

Result on 500 m [s]

[amount]

66 65 64 63 62 61

70 75 80 85 90

Result on 500 m [s]

[amount]

66 65 64 63 62 61

65 70 75 80 85

Result on 500 m [s]

[amount]

66 65 64 63 62 61

250 300 350

Result on 500 m [s]

[amount]

Figure 4: Training loads generated by the RBF network for a hypothetical athlete (age 21, body height 185 cm, weight 75 kg);

The X-axis represents the expected results ranging from 68 s to 62 s. The Y-axis represents the value of training loads.

icSPORTS 2015 - International Congress on Sport Sciences Research and Technology Support

248

The third set of generated training load are runs

over 8–12 hurdles (y

). As can be seen from the pre-

sented graph (Fig. 4(c)) the value of these load in-

creases until the athlete achieves a 64 s result. Sub-

sequently, as the sports level increases so the size of

this load decreases. A similar situation was observed

for y

The ﬁnal set of training loads analyzed are hurdle

runs in varied rhythms (y

). The values of these loads

change in non-linear fashion during the whole period

being considered (Fig. 4(d)). In the early stages a

career the value of these loads is low. It is signiﬁcant

that when the outcome is equal to 65 s the value of y

grows steadily, assuming its maximum value when an

athlete has reached the highest level of ﬁtness.

4 CONCLUSIONS

In this paper the model for generated training loads

to develop techniques was calculated. The model was

calculated using artiﬁcial neural networks with radial

basis functions. The best RBF network has seven neu-

ron in the hidden layer and generates errors at the

level of 21%.

The generated training loads change non-linearly

over the whole of an athlete’s career; the training

loads y

(runs over 8–12 hurdles) can serve here as

an example. Their value increases systematically up

to the moment when the athlete achieves an interme-

diate level (approx. 64 s in a 500 m ﬂat run), and af-

ter that it decreases to the end of the athlete’s career.

The analysis also shows that at a high sports level the

size of y

and y

should be increased (a run over 1–3

and 8–12 hurdles) and the size of y

and y

should be

decreased (runs over 4–7 hurdles and hurdles runs in

varied rhythm).

The implementation of artiﬁcial neural networks

with radial basis functions in training loads analysis

can support the hurdles training process. The results

obtained can be regarded as suggestions to be used

while planning these loads.

REFERENCES

Alejo, B. (1993). Weight training for the 400-m hurdler.

Track Technique, (123):3915–3918.

Bishop, C. M. (2006). Pattern Recognition and Ma-

chine Learning. Information Science and Statistics.

Springer-Verlag New York, Inc., Secaucus, NJ, USA.

Coh, M., Dolenec, A., Toma

zin, K., and Zvan, M. (2008).

Dynamic and kinematic analysis of the hurdle clear-

ance technique. In

Coh, M., editor, Biomechanical

Diagnostic Methods in Athletic Training, pages 109–

116. University of Ljubljana.

Iskra, J. (2012). Scientiﬁc research in hurdle races. AWF

Katowice.

James, G., Witten, D., Hastie, T., and Tibshirani, R. (2013).

An introduction to statistical learning. Springer.

Maszczyk, A., Golas, A., Pietraszewski, P., Roczniok, R.,

Zajac, A., and Stanula, A. (2014). Application of neu-

ral and regression models in sports results prediction.

Procedia-Social and Behavioral Sciences, 117:482–

487.

McFarlane, B. (2000). The Science of Hurdling and Speed.

Athletics Canada.

Perl, J., Grunz, A., and Memmert, D. (2013). Tactics anal-

ysis in soccer–an advanced approach. International

Journal of Computer Science in Sport, 12(2013):33–

44.

Pfeiffer, M. and Perl, J. (2006). Analysis of tactical struc-

tures in team handball by means of artiﬁcial neural

networks. International Journal of Computer Science

in Sport, 5(1):4–14.

Przednowek, K., Iskra, J., and Przednowek, K. H. (2014).

Predictive modeling in 400-metres hurdles races. In

2nd Int. Congress on Sport Sciences Research and

Technology Support - icSPORTS 2014, pages 137–

144. SCITEPRESS, Rome, Italy.

Roczniok, R., Ryguła, I., and Kwa

sniewska, A. (2007). The

use of Kohonen’s neural networks in the recruitment

process for sport swimming. Journal of Human Kinet-

ics, 17:75–88.

Ryguła, I. (2005). Artiﬁcial neural networks as a tool

of modeling of training loads. In Engineering in

Medicine and Biology Society, 2005. IEEE-EMBS

2005. 27th Annual International Conference of the,

pages 2985–2988.

Silva, A. J., Costa, A. M., Oliveira, P. M., Reis, V. M.,

Saavedra, J., Perl, J., Rouboa, A., and Marinho, D. A.

(2007). The use of neural network technology to

model swimming performance. Journal of Sports Sci-

ence and Medicine, 6(1):117–125.

StatSoft, Inc. (2011). STATISTICA (data analysis software

system), version 10. www.statsoft.com.

Wiktorowicz, K., Przednowek, K., Lassota, L., and Krzes-

zowski, T. (2015). Predictive modeling in race walk-

ing. Computational Intelligence and Neuroscience.

Planning Training Loads to Develop Technique and Rhythm in the 400m Hurdles using RBF Network

249