Over fairly recent years the concept of an artificial sport trainer has been proposed in literature. This concept is based on computational intelligence algorithms. In this paper, we try to extend the artificial sports trainer by planning fitness training sessions that are suitable for athletes, especially during idle seasons when no competition takes place (e.g., winter). The bat algorithm was used for planning fitness training sessions and results showed promise for the proposed solution. Future directions for development are also outlined in the paper.
Sport becomes highly addictive for many people in the
world. A few decades ago, people around the world spend
their free time doing different activities, like: short walks
through the park, visiting the cinema or galleries,
fishing, visiting a thermal spa and also meet friends. A lot
of leisure studies also proved this. However, over recent
decades, lifestyles have been substantially changed
especially because of globalization that has transformed the
whole earth into a global village. Due to a lack of time
as well as personal willingness, people do not want to live
like they used to. For this reason, different kind of new
activities have emerged over recent past years. One of
the bigger revitalizations has been sport which has became
extremely in popular because of the emergence of
different mass sports events. For instance, the main mass sport
events are:
• Road marathons - A road marathon [
Participants of the mentioned events participate mostly
because of two goals. The first goal is to enjoy (in other
words: to have a nice time) and the second is to finish the
trial. Usually, finishers are awarded with medals which
are big stimulants for participants. In this case, every year
many more participants have also began to take these
competitions more seriously, i.e., semi professional. In line
with this, they invest much more time in preparations for
competitions. Unfortunately, there is a long way for good
preparations for such kinds of competitions. This good
preparation consists of proper sports training, good
eating and also good resting. To maintain all these factors as
high as possible is very hard for numerous athletes, since
they do not have enough experience. Newbie athletes
especially suffer from the unwanted effects of irregular
training called over-training syndrome [
In order to break this barrier, we began the development
of an artificial sport trainer. An artificial sports trainer
was presented recently in [
This paper, extends the artificial sport trainer with
planning fitness training sessions. These kinds of training
sessions are very important for athletes especially during idle
seasons. In Europe, the idle season is usually during
winter months when the athletes prepare their form for the
whole season. Planning fitness sessions were performed
using a bat algorithm [
Organization of the remainder of this workshop paper is as follows: in section 2 we discuss about characterists of fitness training, while section 3 presents swarm intelligence algorithms and bat algorithm. Experiments and results are presented in section 4, while section 5 concludes the paper. 2
This study, focused on fitness training in regard to cycling.
Incorporation of strength training in cyclists preparatory
periods has received more attention over the last two
decades. Most of the serious and competitive cyclists also
include strength training in their training programs. It is
also evident in some previous research that adding strength
training to an endurance training program can increase
endurance performance [
A combination of endurance and strength training (concurrent training) might therefore be a potential training strategy for promoting muscle oxidative capacity. It might be related to an improved cycling economy, as observed after adding strength training to the ongoing endurance training, namely because a stronger muscles at a certain intensity operates longer with a lower percentage of maximum capacity. It is well-known that adding strength training to endurance training can increase the maximal strengths and rate of force developments in cyclists. In theory, this may improve pedaling characteristics by increasing peak torque in the pedal stroke, reducing time to peak torque and reducing the pedaling torque relative to maximal strength, which in turn may allow for higher power output and/or increased blood flow.
The most important thing for developing a cycling
strength program is to know, which muscle groups are
the most active during the pedal stroke. Some previous
studies have detected a strong correlation between cycling
performance and some strength exercises, like leg presses,
squats, and deadlifts [
For the smart planning of sports training, quantifications, regulating the intensity of a workout is the key for success as indicated as basic knowledge in sports training literature. This fact also holds for fitness training. As an estimate of the intensity of a fitness workout, two main measures are employed like a: • the number of repetitions per set of exercises (NR), • the maximum amount of weight that can be generated in one maximum contraction (1RM).
The logic behind the first measure is as follows. The heavier the weight, the higher the intensity and the fewer repetitions (also reps) an athlete will be able to lift it for. On the other hand, the 1RM determines the desired load for an exercise (typically as a percentage of the 1RM). Let us notice that a coach determines the measure of 1RM for a definite athlete using tests at the beginning of the fitness training and then calculates the number of repeats (NR) in regard to this characteristic value. 3
Swarm intelligence (SI) is a paradigm that belongs to
computational intelligence (CI). According to the [
Algorithm 1 Swarm Intelligence 1: initialize_population_with_random_candidate_particles; 2: eval = evaluate_each_particle; 3: while termination_condition_not_meet do 4: move_particles_towards_the_best_individual; 5: eval += evaluate_each_particle; 6: select_the_best_individuals_for_the_next_generation; 7: end while
Next subsection describes the mentioned algorithm in detail. 3.1
The bat algorithm was developed by Yang in 2010. The
main purposes of this algorithm were to be: simple,
efficient and applicable to varios problem domains. The
inspiration for the bat algorithm came from the phenomenon of
the echolocation characteristics of some types of
microbats. Developer used a three simplified rules describing
the bat behavior, as follows [
The algorithm’s pseudo-code is presented in Algorithm 2.
The main bat algorithm components [
Input: Bat population xi = (xi1, . . . , xiD)T for i = 1 . . . N p, MAX _F E.
Output: The best solution xbest and its corresponding value fmin = min( f (x)).
1: init_bat(); 2: eval = evaluate_the_new_population; 3: fmin = find_the_best_solution(xbest); {initialization} 4: while termination_condition_not_meet do 5: for i = 1 to N p do 6: y = generate_new_solution(xi); 7: if rand(0, 1) > ri then 8: y = improve_the_best_solution(xbest) 9: end if{ local search step } 10: fnew = evaluate_the_new_solution(y); 11: eval = eval + 1; 12: if fnew ≤ fi and N(0, 1) < Ai then 13: xi = y; fi = fnew; 14: end if{ save the best solution conditionally } 15: fmin=find_the_best_solution(xbest); 16: end for 17: end while • evaluate_the_new_solution (line 10): evaluating the new solution, • save_the_best_solution_conditionaly (lines 12-14): saving the new best solution under some probability Ai, • find_the_best_solution (line 15): finding the current best solution.
Generating the new solution is governed by the following equation:
Q(t) = Qmin + (Qmax − Qmin)N(0, 1),
i vi(t+1) = vti + (xti − best)Qi(t), xi(t+1) = xi(t) + vi(t+1), where N(0, 1) is a random number drawn from a Gaussian distribution with zero mean and a standard deviation of one. A RWDE heuristic implemented in the function improve_the_best_solution modifies the current best solution according to the equation:
x(t) = best + ε Ai(t)N(0, 1), where N(0, 1) denotes the random number drawn from a Gaussian distribution with zero mean and a standard deviation of one, ε being the scaling factor, and Ai(t) the loudness.
Contemporary work on bat algorithms captures many
variants and application domains. Some recent works are
presented in papers [
Based on the original bat algorithm, we have developed a modified bat algorithm for planning fitness sessions. Development of this algorithm demanded the following four steps: • determining the fitness exercises, • defining constraints, • modifying the original bat algorithm, • representing the results and their visualizations. In the remainder of this paper, all these steps are described in detail.
Selecting Fitness Exercises. We need to determine specific exercises for different muscle groups before the fitness training can start. Although sports medicine recognizes more than 15 muscle groups that must be included within fitness training, we focus on four groups (i.e., legs, core, arms and back) in this preliminary study. Furthermore, some of these groups can be repeated during the training. Each muscle group is associated with three prescribed exercises as presented in Table 1.
Muscle groups
Exercise LEGS CORE ARMS LEGS BACK CORE LEGS ARMS
LEG PRESS, SQUATS, LUNGE LEG SCISSORS, PLANK, LEG LIFTS PULLDOWN, PUSH UPS, UNDERARM ISOMETRIC EXERCISE LEG PRESS, SQUATS, LUNGE BACK EXTENSION, DEADLIFT, BAR ROWS LEG SCISSORS, PLANK, LEG LIFTS LEG PRESS, SQUATS, LUNGE PULLDOWN, PUSH UPS,
UNDERARM ISOMETRIC EXERCISE On the other hand, the intensities of the exercises must be determined in a fitness training plan. This intensity is associated with a measure 1RM measured for a specific athlete. Here, three levels of intensity are supported in our study, where each level is mapped according to the 1RM, as can be seen in Table 2.
1RM measure 1RM > 80% 60% < 1RM ≤ 80% 1RM ≤ 60%
Note that the data in Tables 1 and 2 were specified according to the suggestions of fitness trainers. Defining constraints The purpose of the fitness training plan is to prescribe sufficient numbers of exercises for each of the prescribed muscle groups, their number of repeats (NR) and the proper intensities (%1RM) such that an athlete simultaneously develops all the muscle groups needed for building the cyclist’s basic form. Therefore, trainers determine the proper amount of a specific exercise in the plan in regarding to the others. In order to regulate the relations between exercises in the fitness training plan, the following constraints are defined: • at least four exercises must have the number of repeats over 25 times (i.e., NR>25), • each training plan should have at least two exercises of high intensity, • each muscle group repeating in Table 1 more than once does not have the same exercise, • if the last exercise in the fitness training plan was of higher intensity, the next exercise should be of medium or high intensity.
In the remainder of this paper, these constraints were
captured within the algorithmic structure of the original
bat algorithm for planning the fitness training plan.
Modifying the original bat algorithm Each solution in
the modified bat (MBA) algorithm consists of 24
floatingpoint elements representing the fitness training plans for
some athlete. The elements of the solution are divided into
three groups of elements. In other words, the solution is
expressed as
xi = (xi1, . . . , xi8, xi9, . . . , xi16, xi17, . . . , xi24)T ,
(3)
where elements xi1, . . . , xi8 denote exercises from Table 1,
xi9, . . . , xi16 are the number of repeats NR and xi17, . . . , xi24
the corresponding intensity, respectively. This means,
each fitness training plan consists of eight exercises with
an assigned number of repeats and corresponding
intensities. While the number of repeats is selected from
interval NR ∈ [
HIGH, if 0 ≤ xi, j < 31 , MEDIUM, if 13 ≤ xi, j < 32 ,
LOW, if 23 ≤ xi, j < 1, respectively. The planning of the fitness training sessions is defined as a constraint satisfaction problem that is formally defined as
Minimize
15 subject to ∑ y j ≥ 4, j=8 24 ∑ z j ≥ 2, j=16
k<4 f (xi) = ∑ χk(xi),
k=1 where and xi,1 6= xi,4 6= xi,7 ∧ xi,2 6= xi,6 ∧ xi,3 6= xi,8, int(xi, j) ≡ HIGH ⇒ int(xi, j+1) 6= HIGH, y j = +1 if xi, j ≥ 25, +0 otherwise, z j = +1 if int(xi, j) ≡ HIGH, +0 otherwise,
The proper solution to the problem is found, when the f (x) = 0.
Representation of Results. Although the results could be visualized, the numerical results in the tables are presented only in this preliminary version of the modified bat algorithm. 4
The results of our experiments are illustrated in Tables 3 to Table 5, where the tables represent the three sets of exercises. An athlete has some free time for resting after finishing each set. We run algorithm 25 times and after the run we selected three generated training sessions which were successfully found by bat algorithm.
LUNGE LEG SCISSORS PULLDOWN SQUATS BAR ROWS LEG LIFTS LEG PRESS PUSH UPS
The obtained results were evaluated by human trainer who evaluated and approved it. The obtained results confirm that the idea of automatic fitness training sessions was worth investigation and the promising results also show the potentials of the solution when used in practice. On the other hand, we would also like to present some problems and bottlenecks which we encountered during development. Firstly, it seems that it will be good to test our idea with evolutionary algorithms in the future. Experiments showed that the success of the bat algorithm in satisfying all constraints was about 25% of runs only. The problem is that the bat algorithm is highly dependent on the best solution. From this reason, our algorithm went into local optimum a lot of times. We believe that advanced mechanisms e.g. arithmetic crossover would behave much better. Moreover, using adaptive and self-adaptive bat variants could also be suggested since we spent a lot of time tuning parameters. On the other hand, many more constraints should be defined in order to have very precise solutions which should be very similar to those solutions created by the human sport trainer. 5
In this workshop paper, we presented a simple, yet efficient solution for planning fitness training sessions automatically. The bat algorithm was employed in order to tackle this problem. This algorithm successfully generated training sessions which were evaluated and confirmed by a human trainer who had more than 20 years of experience. In the future, there are many tasks to do in this direction like for example testing with other nature-inspired algorithms, employing arithmetic crossover and taking more constraints and exercices into account.
The research reported in this paper has been partially supported by the Czech Science Foundation grant 13-17187S.