THE USE OF ARTIFICIAL NEURAL NETWORKS IN SUPPORTING THE ANNUAL TRAINING IN 400 METER HURDLES

This paper presents an evaluation of the annual cycle for 400 m hurdles using artificial neural networks. The analysis included 21 Polish national team hurdlers. In planning the annual cycle, 27 variables were used, where 5 variables describe the competitor and 22 variables represent the training loads. In the presented solution, the task of generating training loads for the assumed result were considered. The neural models were evaluated by cross-validation method. The smallest error was obtained for the radial basis function network with nine neurons in the hidden layer. The performed analysis shows that at each phase of training the structure of training loads is different.


Introduction
The 400 m hurdles race is a complex motor and rhythm (technical) athletics race.In terms of motor preparation, the dominant part is endurance of a specific character (anaerobic), supported by a high level of speed and strength.Given the interdisciplinary nature of race training those means, which combine both the technical and the motor aspects of the race, should be used on a very frequent basis (McFarlane, 2000;Iskra, 2012b).
The analysis of training loads in selected disciplines and sports competitions evokes different reactions among scholars and coaches.Some of them claim that the evaluation of an athlete's (or group of athletes) training can be an inspiration to other sportsmen.Others believe that sport is about individual cases where patterns or "average" data have no value (Hiserman, 2008;Iskra, 2012a).
In the analysis of training loads in athletics, including the 400 m hurdles, three approaches can be distinguished: -Analysis of individual training programme -analysis of the intensity and content of the training of the best competitors, usually record holders and champions (Olympic, world and continent) (Alejo, 1993;Iskra, Widera, 2001;Winckler, 2009).-Statistical analysis of average data -from a group of competitors, who often train over the long term (Brejzer, Wróblewski, Koźmin, 1984;Iskra, 2001;Guex, 2012).-Mathematical analysis -it is an attempt to use basic science to provide training solutions in competitive sports (Iskra, Ryguła, 2001;Przednowek, Iskra, Cieszkowski, Przednowek, 2015;Wiktorowicz, Przednowek, Lassota, Krzeszowski, 2015).Each of the above methods of training load analysis has its strengths and weaknesses.For example, the use of artificial neural networks allows a multidimensional analysis of training loads to be carried out, by creating a system that not only analyses the training already carried out, but also lets the coach decide on the size of the training loads to be applied at a given phase in the sports training.The system which is built on the basis of knowledge accumulated over many years of coaching will assist decision-making by providing valuable coaching tips (Przednowek, Iskra, Cieszkowski, Przednowek, 2015).It should be noted that such a system will act as a consultant, since a coach's intuition and the human capacity to analyse reality is still unsurpassed by computer systems.
The aim of this study is to evaluate the annual preparation cycle for 400 m hurdlers using neural networks.The analyses can be helpful in verifying the views adopted a-priori by coaches, taking into account long-term standards of periodization of training.

Material and methods
The analysis included 21 Polish hurdlers aged 22.25 ±1.96 years participating in competitions from 1989 to 2011.The athletes had a high sport level (the result over 400 m hurdles: 51.26 ±1.24 s).They were the part of the Polish National Athletic Team Association representing Poland at the Olympic Games, World and European Championships in junior, youth and senior age categories.The best result over 400 m hurdles in the examined group was equal to 48.19 s.The collected material allowed for the analysis of 48 annual training plans.
In the presented solution the task of generating training loads (GT) for the assumed result were considered.The neural model generates training for the expected result and the parameters of the athlete (Figure 1 and Table 1).Table 1 contains the variables considered and their basic statistics, i.e. the arithmetic mean of x, the minimum value x min , the maximum value x max , standard deviation SD and the coefficient of variation V.This study uses artificial neural networks in the form of the multilayer perceptron (MLP) and the radial basis function (RBF).Multilayer perceptron is the most common type of artificial neural networks (Bishop, 2006).During MLP training, exponential and hyperbolic tangent function were used as the activation functions of hidden neurons.The feature of RBF network is the fact that the hidden neuron performs as a basis function that changes radially around the selected center.All the analysed networks have one hidden layer.For the implementation of neural networks, StatSoft STATISTICA software was used (Statsoft, 2011).The cross-validation method was implemented using Visual Basic language.
The models presented in this paper were evaluated by leave-one-out cross-validation (LOOCV) (Arlot, Celisse, 2010).The idea of this method is based on the separation from data set n subsets, where n is the number of all patterns.Each subset is formed by removing from the data set only one pair, which becomes the testing pair.The cross validation error (CVE) is expressed by the formula: where: NRMSE j -the normalized root mean square error for the j-th output, r -the number of outputs, n = 48 -the number of patterns, y ij -the real (measured) value, ŷ -ij -the output value constructed in the i-th step of crossvalidation based on a data set containing no testing pair (x i , y i ), y jmax -the maximal value of the j-th training load, y jmin -the minimal value of the j-th training load.

Results with discussion
The main aspect of supporting sport training presented in this study is generating training loads for selected parameters of an athlete.In this way, the proposed approach allows, among others, for individualization of a training plan (Bompa, Haff, 1999).
Taking into consideration various topologies of networks, an optimal multi-layer perceptron was calculated.This model has 5 neurons in the hidden layer and hyperbolic tangent activation function.Compared to the best model with an exponential function it is superior because it generates the error smaller by 0.2%.The best perceptron generates the annual training plan with the error CVE = 19.95%(Figure 2).The optimal RBF network has 5 hidden neurons and CVE = 19.34%.This result is better than that obtained for the MLP networks.Therefore, as the optimal method, the RBF network with five hidden neurons was used.The optimal model was analysed to determine the errors generated for different outputs, which allowed to identify which training means are generated with the smallest error (Table 2).The detailed analysis showed that y 4 , y 9 and y 14 Vol. 17, No. 1/2017 The Use of Artificial Neural Networks in Supporting the Annual Training in 400 Meter Hurdles (speed endurance, strength endurance II, upper body strength) are generated with the highest accuracy (NRMSE j at the level of 14-15%), whereas the output representing technical exercises in march (y 17 ) has the largest error (30%).The chosen neural network was tested by generating training plans for a hypothetical athlete (age: 21 years, body height: 185 cm, weight 75 kg).In every case the result was expected to improve by one second as a result of accepting the output from 56 to 49 seconds.Training loads forming speed (y 1 -y 3 ) are very similar in nature (Figure 3).At the beginning of an athlete's career, the highest content of these loads can be noted, and with their increasing competitive level, a decrease in loads (until the competitor achieves 51 s), can be observed.While obtaining the best results, the rates of training loads influencing speed go up with increasing sports level.
The 400 m hurdles race is still a sprint distance so the need for speed training is the priority, but requires a variety of assessments in terms of a year-round and long-term cycle of preparation.For "high-speed" hurdlers short races can be an important part of training, but in the group of other hurdlers ("endurance" and "rhythm" type) maximum speed exercises are only additional to the basic training (Iskra, 2012b;Balsalobre-Fernández, Tejero-González, del Campo-Vecino, Alonso-Curiel, 2013).Analyses show a characteristic tendency to reduce the importance of speed training in the middle phase of the development of a sports career with a return to speed exercises for the highest performance (y 1 -y 3 ).This fact can be explained, on the one hand, by a particular emphasis on anaerobic exercise during the period of "growing up" to athletic championship level, and, on the other hand, by shortening distance of training at the final, highest phase.Such tendencies can be observed in the analysis of the content of training of the best Polish hurdlers who have been competing for many years (Iskra, Widera, 2001).
In the group of endurance training loads (y 4 -y 7 ) two trends of changes depending on the level of training (Figure 3) were observed.The content of exercises that form speed endurance (y 4 ) and aerobic endurance (y 7 ) increases when the athlete obtains average results (up to 52-51 s), while at a later phase, when his/her form is improving, the value of these loads is consistently declining.Other training loads related to strength have similar tendencies to the speed loads.The values of these loads (y 5 , y 6 ) at the beginning gradually decrease until the competitor achieves 52 s.The values start rising with the increasing competitive level of the athlete.
The whole essence of the running training of a 400 m hurdler, supported by research in the physiology of physical effort, lies in the statement above.The 400 m hurdles distance is a typical anaerobic effort, for which the value of lactate amounts to 20 mmol/l (Ward- Smith, 1997;Gupta, Goswami, Mukhopadhyay, 1999;Zouhal et al., 2010).Therefore, the best, in terms of motor skills, competitors use specific training means at the prime time of their career.Including "alternative" sets of exercises of reduced intensity in this period (the so-called "tempo endurance" system) can be explained by the difficult conditions for Polish winter training, which encourages coaches to reduce the speed of races in favor of training intensity (Iskra, Przednowek, 2016).Changes in the content of strength and speed exercises of lower and upper limbs (y 15 , y 16 ) have similar variability (Figure 4).At first the changes are very small and the level of the loads is relatively small.The content of these loads is increasingly going up only when the competitor achieves results of 52 s.At the championship level the loads stabilize at a high level.
Improvement of the strength capacity in athletics speed races is now one of the trends in searching opportunities to improve results.It is mentioned by the classics of the theory of sports training (Bompa, Haff, 1999;Sozański, Sadowski, Czerwinski, 2015) and the best coaches of this sport (Smith, 2005;Husbands, 2013).The results of the analyses in the group of the best Polish hurdlers do not confirm entirely this trend.Only the basic strength training exercises of the lower limbs from the "average" level remain at the same high level (y 10 -y 11 ).Attention should be

Figure 1 .
Figure 1.Block diagram of the model for generating training loads

Figure 2 .
Figure 2. CVE error for artificial neural networks

Figure 3 .
Figure 3. Training loads y 1 -y 14 generated for results from 55 s to 48 s

Table 1 .
The variables and their basic statis

Table 2 .
Errors for the outputs of the RBF network