Machine Learning for Individualized Training Support

in Marathon Running

Ladislav Havaš

, Vladimir Medved

and Zoran Skočir

Department of Electrical Engineering, University North, Varaždin, Croatia

Faculty of Kinesiology, University of Zagreb, Zagreb, Croatia

Faculty of Electrical Engineering and Computing, University of Zagreb, Zagreb, Croatia

1 OBJECTIVES

Modern technology significantly influences the field

of human physical exercise monitoring and control.

Assuming a cybernetic-like approach to sports

training, we witness a dynamic, synergistic

interaction of the man-technology system realizing a

training process. In this paper we focus to actual

implementation of Information and Communications

Technology (ICT) for intended improvement of

marathon runners’ training. While general principles

of sports training are known (Milanović, 2009), ICT

offers profoundly novel and original possibilities to

influence and actively control training process of a

particular individual. First author’s sports experience

(Havaš and Vlahek, 2006) is combined with an

approach based on telecommunication platform

which was gradually built, upgraded and validated

over the years (Havaš, at al., 2013) to produce an

original, comprehensive and intelligent system

suited to individualized use (Havaš, 2014).

Here we focus in particular to machine learning

features of the approach enabling a flexible, on line

physiologically based-monitoring training support

system for marathon running. Developed, tested and

validated on marathon runners’ data, the system

however posesses capabilities for application in

sports training in general.

2 METHODS

Data mining and On Line Analytical Processing

(OLAP) analyses are used, autonomously or on

demand. Realized and „imagined“ trainings are

validated and user progress is analyzed so that signs

of overtraining, undertraining or possible disease are

recognized. Using available telecommunication

channels a message is automatically generated and

sent to user and new (corrected) program is created

for remaining number of days of the training

process.

Among many available methods and techniques of

data mining, concepts of neural networks are used

(Neural_network, available at: http://en.wikipedia.

org/Neural_network, accessed on 02 December

2012, Neural_network_theory, available at:

http://fann.sourceforge.net/report, accessed on 05

December 2012). Program modules are developed

for analysis of runners that have their active

programs, and offer to them possible corrections,

periodically, after their registration to the system, or

based on special request by each user. Combining

methods and concepts of data mining and OLAP

analysis (Online_analytical_processing, available at:

http://en.wikipedia.org/wiki, accessed on 14

November 2012), a new method of validation is

developed, defined and realized for trainings of long

distance runners (Havaš, 2014).

Figure 1: Searching for similar profiles.

Havaš L., Medved V. and Sko

cir Z..

Machine Learning for Individualized Training Support in Marathon Running.

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

Periodic control trainings are envisaged based on

which one can monitor and calculate new values of

maximum aerobic capacity VO2

max

. Each positive

(or negative) change is dynamically in calculated in

all remaining training sessions, by means of which

one attains improvement in remaining preparation

period.

Developed prototype searches for similar runner

profile patterns, and based on known history of their

accomplishments forms and displays corresponding

corrections, generating individualized program for

preparation (Figure 1).

By filling data warehouse for even greater

sample of runners in long periods, conditions are

created for reliable prediction of future realizations

and timely correction of preparation program.

3 RESULTS

The majority of tested marathon runners train many

years and have achieved results in marathon, which

were compared with the results that were achieved

with the assistance of developed Online Running

Trainer (ORT) system. The comparison between

previous personal bests (achieved in the past three

years) and those achieved by using the implemented

system is shown in Figure 2.

Figure 2: The comparison between previous and current

results of tested users (in Croatian language).

X-axis of the Figure 2 represents tested users (58

marathon runners), while the y-axis represents the

achieved results of those runners in seconds. The red

curve in the Figure 2 shows previous results in

marathon, while the blue curve shows achieved

results with the ORT system. It can be seen that

almost all tested users have improved their results

and have successfully finished the training cycle

without sport injuries. It is important to mention that

they have not reached that result in several iterations

(training cycles), but within one season in which

they have been systematically guided.

4 DISCUSSION

Mentioned verification process has its flaws and was

not carried out in a way which would (possibly)

show more (or less) deviations from the anticipated

results. Only one group of athletes (58 athletes) was

tested, who used the developed system and ran on

their key races. The results of those runners in the

past three years, achieved without the help from the

developed prototype were compared with the results

with the assistance of ORT system, where there is

only partially possible to compare the parameters of

“different” groups of users who train in different

ways. Because of the specificity of testing sport

achievements, the comparison between the group of

marathon runners and the group of people who do

not actively engage in sport activities (running)

would not have any practical value. Listed

limitations of the conducted research should result in

additional evaluations, for the purpose of obtaining

quality and more certain judgments on a larger

sample of athletes. Periodic loading of data

warehouse (ETL process) will, in time, enable the

evaluation of algorithms and methods on a larger

data sample, while the developed ORT system will,

by the method of self-learning, correct the training

elements and generate better quality programs.

5 CONCLUSIONS

Machine learning means a computing method

enabling future actions be improved based on

information from past (Gamberger, 2011). Presented

method, further, is based not only on the study of

realization by one user, but the system learns as a

whole based on registered data on remaining

runners. This demonstrates that presented method

successfully implements machine learning paradigm

to the problem of improving sports training.

REFERENCES

Milanović, D., 2009., Theory and methodology of

trainings (in Croatian), Faculty of Kinesiology,

University of Zagreb, Zagreb.

Havaš, L., Vlahek P., 2006. Road running (in Croatian),

TK Međimurje. Čakovec.

Havaš, L., Skočir, Z., Medved, V., 2013., Modeling of the

athlete's training decision support, Vol. 20, pp. 315-

322, March-April, Technical Gazette.

Havaš, L., 2014., Information system for athlete's training

decision support (in Croatian), Doctoral Thesis,

Faculty of Electrical Engineering and Computing,

University of Zagreb, Zagreb

Havaš, L., Medved, V., Skočir, Z., 2013, Application of

mobile technologies in the preparations for long

distance running, icSPORTS 2013, Science and

Technology Publications, pp.154-161., Lda,

SCITEPRESS.

Neural_network, http://en.wikipedia.org/wiki

Neural_network_theory, http://fann.sourceforge.net/report

Online_analytical_processing,http://en.wikipedia.org/wiki

Gamberger, D., 2011, Knowledge discovery by data

mining (in Croatian), Institute Ruđer Bošković.