Dynamic Difficulty Adjustment using

performance and affective data in a Platform

Game

Marcos P. C. Rosa1 , Eduardo A. dos Santos1 , Iago L. R. de Moraes1 , Tiago B.

P. e Silva1 , Mauricio M. Sarmet2 , Carla D. Castanho1 , and Ricardo P. Jacobi1
1
University of Brası́lia, Brası́lia, Brazil
2
Federal Institute of Education, Science and Technology of Paraı́ba, Paraı́ba, Brazil

Abstract. The Dynamic Difficulty Adjustment (DDA) of games can

play an important role in increasing the player engagement and fun.
Gameplay difficulty can be adapted according to the player’s perfor-
mance, its affective state or by using a hybrid model that combines both
approaches. This work investigates a hybrid DDA mechanism for a plat-
form game to appropriately adapt its difficulty level and keep the player
in a state of flow. The three approaches are compared to verify the ef-
ficiency of each model. An open source platform game was adapted to
support the hybrid DDA algorithms. Game telemetry was introduced to
acquire performance data and the affective state of the player is esti-
mated through physiological data obtained from the Electrodermal Ac-
tivity (EDA) of the skin. A method that estimates game difficulty when
varying platform size and jump height was developed to support the
DDA process. Besides playing with the different DDA models, each par-
ticipant answered questionnaires and had their data collected for inquiry
purposes. The results indicate that the DDA models were able to adjust
the gameplay difficulty to the players, increasing the number of com-
pleted levels and reducing the variation in the playing time.

Keywords: platform game · dynamic difficult adjustment · affective

game · game telemetry · flow

1 Introduction

The video game industry has seen huge growth over the last few decades, with
game developers appealing to a broad audience to sell their products. Thereby,
the rules and requirements to make good games need to fit new scenarios, cre-
ating fun games for each consumer profile.
Adaptive systems strive to achieve the balance of the game for different
players in a consistent, fair and fun way [17]. Such systems provide a person-
alized experience, adjusting the game according to identifiable, measurable and
influential goals [16]. They are usually defined by a dynamic factor and make
adjustments based on those goals [3].
2 M. Rosa et al.

Non-adaptive games can cause boredom or frustration by not evaluating the

match between the degree of challenge and the player profile [7]. For this, the
activity can keep the player in a state of flow, offering a continuous challenge
according to the evolution of the player’s skills, so that he/she feels motivated
to continue playing with a challenge proportional to his/her performance [18].
The concept of flow offers a theoretical explanation for pleasure [10]. During
the state of flow, attention is free to focus on achieving the player goals as there
is no clutter to strengthen or threat to defend. Thus, the flow occurs when there
is a balance between the abilities of the individual and the undergoing task.
Each player has an experience curve obtained by playing a segment of the game
while remaining in a safe zone. If the experience moves away from the flow zone,
negative psychic entropy (anxiety, boredom) will interrupt the flow state.
In contrast to the pre-established difficulty levels (commonly set as easy, nor-
mal, and hard), Dynamic Difficulty Adjustment (DDA) is an emerging technique
in electronic games that seeks to adapt the level of difficulty of a game while to
suit the skills of the player during the gameplay [1]. There are three requirements
for dynamic game adaptivity [1]:

1. Identify and adapt to the player level as fast as possible;

2. Perceive and record changes in player performance; and
3. Maintain the behavior of the game, during the adaptation, credible and
discreet, so that the player does not perceive the adaptive system.

In short, DDA modifies, in runtime, game factors and variables to match the
challenge to the player. It should act as a control mesh that monitors player and
game variables and acts accordingly, adapting avatar attributes, NPCs or level
variables.
The DDA can be based on the player’s performance data, on the player’s
physiological data or on the combination of these, thus characterizing a hybrid
DDA [5]. Most studies are related to the player’s perceived difficulty, where the
proposed ADD models are based on performance data [11][12] or on the player’s
physiological responses [13][23][6]. Besides, there is scant in-depth research into
the efficiency of using sensors and hybrid adjustments to make the player feel in
a state of flow, especially when considering the platform game genre.
In this work, a hybrid DDA model is implemented and evaluated for a plat-
form game to verify if the combined approach can provide a better difficulty
adjustment for the player. Also, this work aims to compare the state of flow
with the adaptation of distinct variables of the game, such as platforms or the
jump height. Thus, the study is divided into three stages of analysis, comparing
(1) the difficulty estimated by algorithms and the difficulty experienced, (2) the
DDA models, and (3) the performance model applied to distinct game variables.
The adaptations are applied to a 2D platform game, which has a player-
controlled character, called an avatar, that runs and jumps to avoid obstacles
and/or defeat enemies. Based on the most common mechanics, there are 3 main
aspects to analyze platform games [19]:
 Title Suppressed Due to Excessive Length 3

1. Movement, which expresses the range of movements included in the avatar

and the respective control over those movements;
2. Confrontation, which expresses the importance given to an environment with
opponents; and
3. Interaction with the environment, representing additional gameplay features
that are not directly related to the original idea of the game.

According to the proposed objectives, the game adaptation prioritises move-

ment rather than confrontation or interaction with the environment. Thereby,
the game will adapt platforms and player’s jump height configuration through
a hybrid DDA system that considers the player’s performance correlated to an
assessment of their affective state. The difficulty measurement of a component
will be based on the mathematical equations proposed by Mourato [21], related
to the success in executing jumps. Thus, the probability of success or failure of
a jump will determine the relative difficulty.
Furthermore, a data acquisition device (E4 Wristband) is used to collect
physiological data from the player. This sensor checks the skin conductance that
provides a reference for interpreting the level of frustration or boredom of the
player. This type of measure (Electrodermal Activity (EDA)) serves as input
to the DDA system and is usually related to some sympathetic nervous system
reaction, indicating a change in the level of enthusiasm [22].
In each of the stages of analysis defined, tests were made with volunteers
and the flow state and adequacy of the difficulty were evaluated, according to
the answers to the questionnaire. In addition, these data were compared to the
performance of the player according to game variables captured during the test
sessions.

2 Related Work

Araújo conducted a study on player modeling and a survey on comercial and

academic games that use DDA concepts [3], both commercial and academic.
With a 2D shooter, it was analyzed the influence of adaptability on the gaming
experience from the perspective of the flow theory [10]. In this case, the en-
emy’s variables (speed, shooting time and range) are modified according to the
difficulty, with a finite number of lives for the player and pre-defined difficulties.
The work found that changes that are easier to perceive impact the player’s
experience negatively. It has been observed that a poor performer may feel un-
motivated when changes in the game are evident or when he understands how the
adjustment system works. In addition, it becomes possible to break immersion
and abuse the system by the player.
On the other hand, the difficulty in automatically generated 2D platform
games was the research theme of Mourato and Santos [21]. For this, a method of
measuring difficulty was proposed from the probability of success to complete a
level. What they considered an obstacle was simplified and it was not considered
the presence of checkpoints or lives. Moreover, they used the concept of levels
4 M. Rosa et al.

as a path between a start and an end point that may contain components that
hurt the player [9].
The measurement of a level difficulty was examined by the probability of the
player giving up, repeating the challenge and succeeding in a jump. Besides, fail-
ure can occur when taking damage and dying or becoming frustrated and giving
up. In turn, the probability of success of the jump is estimated by measuring
the difficulty introduced by each individual component, based on formulas from
kinematic physics.
Therefore, the combination of the probabilities of success of each jump situ-
ation produces an estimation of the level of difficulty. The height and position of
a platform, the distance between platforms and the position of the enemies can
be parameters that determine the success of the jump and facilitate or hinder
the level.
Following the studies carried out in [21] and [20], Mourato discussed the
improvement of Procedural Level Generation (PLG) in platform games [19].
Content adaptation is guided by the user, analyzing patterns of success and
failure. Thus, the proposed set of difficulty estimation metrics is used to increase
the capacity of the PLG algorithms and to customize the experiment. From this,
it has implemented an algorithm for generating global structures and another for
adapting the content, transforming simple paths into challenges with complex
structures when changing components in strategic locations.
Difficulty can be adapted according to the player’s performance, its affec-
tive state or the combination of both models. Considering this scenario, a facial
expression analysis was applied as a means to predict difficulty through a classifi-
cation task [4]. In this sense, it was needed to define metrics to measure perceived
difficulty through affective data analysis, as well as model player behaviour and
predict their affective state from this data.
The study was able to predict difficulty by correlating facial expressions to
distinct player emotions. Despite that, these methods need a wider range of
testing and improvements in accuracy. Also, they may evolve by integrating
performance and physiological data, such as the Electroencephalography (EEG)
and the Electrodermal Activity (EDA) of the skin.

3 Approach for Dynamic Difficulty Adjustment

The present work deals with the development of a hybrid Dynamic Difficulty
Adjustment system for platform games. Since the purpose is not to develop a
new electronic game, efforts were focused on adapting an existing game to meet
the needs of the research. Thus, the platform game The Explorer: 2D, developed
by Unity Technologies, was chosen, where the avatar can move horizontally,
jump, crouch, attack and shoot.
The ready-made game levels have enemies of different ranges and non-linear
paths, making it necessary to solve puzzles such as locating collectibles or stand-
ing on pressure panels to open doors. Regarding this study, among the three main
components of the platform genre [19], the focus is on movement rather than
 Title Suppressed Due to Excessive Length 5

confrontation and interaction with the environment. In this way, the game was
adapted for level generation by bitmaps. Also, when creating levels, platforms
and player variables are prioritized, in view of their more deterministic nature
[14], such as in a classic platform game.
All of the game components were categorized, as the avatar, end point, check-
points, collectibles, platforms, destructible or pushable objects, death areas, and
enemies. Each level is organized into a map of components through the bitmap
input, with each component changeable by applying the DDA mechanism.

3.1 Data extraction

Gameplay and player data are needed for Dynamic Difficulty Adjustment and
analysis. Performance data was captured directly from the game and real-time
physiological data was acquired with the Empatica E4 wristband, a wearable
sensor device. The Electrodermal Activity (EDA) signal was chosen due to its
correlation with the levels of excitation, tension and frustration of the player.
This type of sensor enables DDA to achieve immersion [15], and the reliability
of this type of measure has been tested in previous studies [22].
Empatica Inc. provides Android and iOS applications designed to enable real-
time data flow from the E4 devices, and a Bluetooth connection with cell phones.
As a result, the default developer application has been modified to send the raw
data to a database server over an internal network. Also, the game was adapted
to request the recent data orderly, with the raw data of the EDA processed to
generate two derived data: tonic (underlying slow changes) and phasic (rapidly
changing peaks).
On the other hand, the gameplay data was organized into time-related, quan-
titative and death-based data. These are, respectively, the time to end each level,
the quantity of successful jumps, and the distance travelled before dying. Fur-
thermore, the success of a jump is related to two main goals established for the
player: respectively, to reach the end of the level, and to obtain collectibles.
Considering these goals, the type of platform is divided according to the
expected origin and destination of the avatar in relation to the position of the
platform. In this sense, each platform carries out informations such as if the
player must traverse from left or right, and from below, above or on the same
height, always aiming to reach the end of the level. From this, the jumps are
categorized into successful, flawed or randomic when considering the factors:

1. Fall into an chasm or die through an enemy or a trap;

2. Change of platform, considering the origin, the destination, the distance
traveled and the collisions of the avatar;
3. Surpass a gap or an chasm;
4. Get collectible.

With this type of data, it is possible to measure the player’s performance

while traveling a level, regarding the jumps that are made and the consequent
interactions with platforms, gaps, collectibles, and death areas.
6 M. Rosa et al.

3.2 Difficulty measuring

To evaluate the difficulty of a level in a platform genre game, this work im-
plements a version of the Mourato method [21][19]. His method evaluates the
generation of levels and the adaptation to generate more complex and non-linear
structures, such as gates and levers. Thus, it performs a procedural content gen-
eration associated with the variation of the environment complexity, rather than
an adjustment of the difficulty of the components of the game.
In relation to the representation of difficulty, it follows the idea of Aponte [2]
which is related to the probabilities of success and failure during the execution of
tasks, whether analyzed independently or at an entire level. Thus, the approach
is based on two steps of level analysis:

1. First, the level is decomposed into segments that represent independent parts
of a level, with possible transitions between them.
2. Then an analysis is made of each individual component, based on its prob-
ability of success.

To determine the difficulty of a level, the sum of the difficulty of specific

components is made based on its probability of success. The analyzed compo-
nents are those that have jump as an integral part of their challenge, being these:
platforms, gaps, and chasms. Moreover, the jump corresponds to a parabola of
the launch of a projectile from an origin to a destination platform, with the
probability of success coming from the horizontal and vertical margin error of
the parabola in relation to the destination point.
For moving platforms, the failure is defined when the user collides with the
entity in the period of loss or if its movement is late and does not fit into the
entity’s overcoming interval. Combining the two, the spatiotemporal difficulty
is measured by the multiplication of the success probabilities of the previously
mentioned situations.
To automate this process at runtime, for each platform component in the
segment, a search is made to find the shortest path from the origin platform (P0 )
to a destination platform (P1 ). There must always be a reachable destination
when adapting, according to the desired movement stored in the component.

3.3 Dynamic Difficulty Adjustment

The DDA of the game is adjusted: (i) according to factors of player performance,
(ii) another by the affective state evaluation of the EDA and (iii) a hybrid of both.
In the three models described, the difficulty is changed in the same way. In short,
there are global values that measure the player’s performance and the affective
state, varying from 0 to 1. The difficulty for the hybrid model corresponds to
the average of the others.
The implementation of the three DDA models followed a cyclic pattern with 4
interconnected systems, based on Chen [8] and Bontchev [5] studies. The systems
refer to the Player’s raw data, the Monitoring of these metrics, the corresponding
 Title Suppressed Due to Excessive Length 7

Analysis and the Control of the components that will be adjusted. The Figure
1 shows the diagram of the systems used to apply hybrid DDA and a summary
of the associated steps.

Fig. 1. Cyclic diagram of the system-driven hybrid DDA. The systems are in green,
the steps related to the player’s performance in red and the affectivity in yellow.

The Analysis System applies heuristics for both performance and affective
data, generating a global difficulty value used to define the desirable mean prob-
ability of jump success. For the affective case, the number of arousal levels is ex-
tracted based on the normalization of individual peaks by the asset from Project
Rage3 . At each time period, the ratio is recalculated and its variation is added
to the overall difficulty.
As for the performance case, heuristics were defined. When death is related
to a fall, a counter is incremented and the position of death is saved. If death
is in a previous or equal position to that of the current death, the overall diffi-
culty decreases (facilitating). In the opposite case, the global difficulty increases
(hampering), as the player was able to further advance in the level.
When a segment is completed and at least one death occurs, the difficulty
decreases. Meanwhile, the difficulty grows when segments are completed without
the occurrence of death. Finally, after each death or segment completion, a ratio
is made between the number of successful and fault jumps. Thus, it makes harder
when there is a greater incidence of success, and facilitates when the number of
failures increases.
The DDA adaptations were programmed to be made when creating segments
and levels, and when dying. It is observed that they are not done at every moment
to avoid that the player perceives the visual modifications. Moreover, to maintain
an increasing difficulty throughout the game, the inferior and superior global
3
https://github.com/ddessy/RealTimeArousalDetectionUsingGSR
8 M. Rosa et al.

difficulty limits are defined according to the number of levels in the game. For
example, it must be between 0 and 0.75 for the first level, and 0.25 and 1 for the
last level.
The Control System is responsible for applying the DDA to the game and
player variables. Thus, on the x-axis, the size of gaps and platforms is modified,
as seen in Figure 2. Associated to the size of the gaps, it adds (by facilitating)
or removes (in hindering) tiles from the region of destination of the jump.

Fig. 2. Segment screenshots demonstrating the difference for the horizontal platform
distance (a) without the application of DDA, and with DDA making the level (b) easier
and (c) harder, respectively.

On the y-axis, the position or size of the platforms is changed, shown in 3. In

the case of column-shaped platforms, it is changed by removing or adding tiles,
otherwise the position of the platform that is modified. Moreover, the player’s
jump amplitude was adapted in parallel or together with game variables in a
composite Control System. As seen in Figure 4, the amplitude can be increased
to make it easier to access more distant platforms.

Fig. 3. Segment screenshots demonstrating the difference for the vertical platform dis-
tance (a) without the application of DDA, and with DDA making the level (b) easier
and (c) harder, respectively.

At the end of the process, the difficulty of each component is updated, as

well as the level difficulty. Note that the Control System of the DDA adjust the
game variables and the player configuration guaranteeing that all game content
remain reachable during adaptation.
 Title Suppressed Due to Excessive Length 9

Fig. 4. Segment screenshots demonstrating the difference for the jump amplitude (a)
without the application of DDA, and with DDA making the level (b) easier and (c)
harder, respectively.

4 Experiments

Three test batteries were made with sample groups. In each battery, the player
could pause, review the controls and give up on a level. Moreover, lives were
infinite and levels were divided into segments of same length delimited by check-
points, in which the progress was saved. The game session was finalized at the
end of all levels and when the questionnaire was fully answered.
In the first battery, the game presented 12 levels, each containing one or two
segments. The difficulty during the game was increasing and analyzed, in each
level, different components of the game (or the conjunction of these). It was
expected to determine a balance of the difficulty of the game and to analyze the
method of estimation of difficulty proposed and implemented in the platform
game The Explorer: 2D. Therefore, the objective was to verify if the estimated
difficulty is able to relate to the difficulty experienced by the players.
In the second battery, the game featured 12 levels, each containing four seg-
ments with similar lengths. It followed the pattern: (1) level without the appli-
cation of DDA, (2) with DDA based on player performance, (3) with affective
DDA, and (4) with hybrid DDA. This pattern was repeated 3 times, whereas the
components and base difficulties were similar in each set of levels, the adjust-
ment could change the general difficulty according to the DDA model applied in
the level.
The four cases were compared and similar components and obstacles in each
set were selected in order to avoid particular factors biasing the conclusions, such
as the usage of different game components or exploratory paths. From this, the
following hypothesis was defined:

Theorem 1. A hybrid DDA system with performance and affective data can
provide suitable difficulty for the player.
10 M. Rosa et al.

In the third battery, the game also featured 12 levels, simplifying the general
presentation of those used in the second battery to allow remote tests. In this
case, a Control System adaptation was randomized for each volunteer, and it
was guaranteed that each played only once. The available adaptations were:
(1) game without the application of DDA, (2) with performance-based platform
adaptation, (3) with performace-based jump height adaptation, and (4) with
performance-based combined adaptation of both platform and jump height. In
this sense, all the levels were played entirely by each volunteer with a single
version of the Control System adaptation. Also, the following hypothesis were
defined:

Theorem 2. A combined DDA system with platform and jump height adapta-
tion maintains the player in a state of flow.

Theorem 3. A combined DDA system with platform and jump height adapta-
tion has a suitable difficulty for the player.

The tests were validated by the application of questionnaires during the game
session. The questions applied before the start of the game concerned sociode-
mographic data. During the game, they checked the difficulty perceived at each
level. In the end, they analyzed the perceived difficulty, and whether the player
remained in a state of flow. For the first battery it was also observed what
influences the player’s performance and the challenge provided by different com-
ponents. Meanwhile, for the third battery, there was no in-game questionnaires,
so that the flow state was not harmed.
In all test cases, the performance data were collected and saved. The test
conditions ensured that all participants had the same game information, and,
for the non-remote first and second batteries, the same environment with no
external contact.

5 Discussion

In the first test battery, 20 university students participated as volunteers, ranging

from 18 to 24 years old, with a higher male participation (70%). The players had
a predilection for games with easy difficulty (45%) rather than normal (30%) or
hard (25%).
In the second test battery, the same university context was maintained and
36 volunteers participated, ranging in age from 18 to 25 years old, and with a
greater participation of men (61.1%). There was a predilection for easy difficulty
(50%) than normal (33.3%) or hard (16.7%). Furthermore, 16 of the 20 partici-
pants of the first battery participated again, noting that new levels were created
and several DDA models were applied that should adjust the game regardles of
whether the player knows the game dynamics.
In the third test battery, 155 volunteers participated remotely, ranging in age
from 15 to 65 years old, and with a higher participation of men (81.69%). 12
of the 155 volunteers were discarded for not completing the game session, and
 Title Suppressed Due to Excessive Length 11

11 for being categorized as an outlier, finishing levels with less than 10 seconds
and with one death at most, stating that the player was not committed to the
attempt. The players only had one DDA adaptation applied (37 without DDA,
29 with platform DDA, 35 with jump DDA, and 31 with combined DDA), having
a total of 132 volunteers considered. Also, there was a predilection for normal
difficulty (51.41%) than hard (32.39%) or easy (16.20%).

5.1 Balancing and checking difficulty measurement

The difficulty of each level was numerically described by the participants in a

questionnaire after playing it in the first battery, according to Figure 5. The levels
tested distinct components and their interactions, and it was possible to identify
the proportionality between the normalized value calculated by the algorithm
and the one obtained by the players’ average, between 0 and 10. Therefore,
this estimation method can be used in the application of DDA. However, level
6 showed that the game design made the difficulty to not vary in agreement.
In level 6, the challenge is created through the proposed design, whereas the
algorithm only counts the positions and specifications of the components in the
construction of the difficulty of the levels.

Fig. 5. Comparison, in each level of the first battery, between the average of the difficul-
ties measured by the questionnaire and the value estimated by the developed algorithm.

There was a peak in levels 5, 9 and 12 for having checkpoints and double the
standard size, thus bringing greater difficulty for the accumulation of challenges
and components. In this way, it was possible to corroborate the relationship
between the amount of components and the difficulty perceived by the players.
Besides, greater difficulty has been observed in separate components working
together, and that the modification of specific components alters the user’s per-
ception of difficulty, such as the complementary levels 2 and 7 that alter the
12 M. Rosa et al.

difficulty only by increasing the size of gaps. Therefore, it was decided that the
levels should have the same size and do not present challenge based on the game
design, as the blocking of the path by a riddle.
At the end of the game session, it was questioned about the actions that
most influenced the performance of the players, with the highest goal being to
complete the level (with 80% of frequency among the participants) as opposed
to completing the levels quickly (25%). In addition, there was a low concern in
exploring (30%), and defeating the maximum of enemies (30%). Thus, the DDA
was based on more general components (such as platforms and gaps) and was
tested with linear paths.
Additionally, it was found that different characteristics of a platform influence
the perception of challenge by the participants in a similar way. The volunteers
were questioned about the difficulty that different characteristics of a platform
can provide, such as size and distance, with all getting an average between 3.8
and 3.9 on a scale of 1 to 5. From this, it was decided that the DDA would adapt
both the size and the vertical and horizontal distances of platforms, gaps and
chasms.
Likewise, the moving platforms had similar difficulty considering the direction
of movement (horizontal, vertical or diagonal), and components with the same
purpose and different visual provided similar difficulty, such as the gap and the
acid (average of 3.9). On the other hand, enemies with distance attacks and traps
increased the difficulty (mean of 4.05 and 4.4, respectively), avoiding the use of
these in the construction of DDA analysis levels, since they were not considered
in the adaptation.
The basic levels (as seen without the DDA application) were divided in three
stages of 4 levels for the following tests, according to the overall component
experienced difficulty. The components were divided into easy, normal, and hard
level sets, with enemies with long distance attacks and traps in the hard set.
Note that the destination space of platforms and the size of gaps is modified by
the DDA algorithm, allowing to adapt to the player even if its not in a specific
component difficulty. In this sense, initially all levels of future tests need to have
each set with similar estimated difficulties, and the same types of components
organized in different ways to create variation.
There was a high rate of participants completing the levels, with a decrease
according to the greater difficulty of the levels, and a high variation in the dwell
time at each level, especially when analyzing levels with hard components.It was
observed that the possibility of giving up a level was chosen after many deaths
or with a high level of frustration of the participant at the level. These factors
are intended to be normalized with the presence of the DDA and by avoiding
nonlinear paths.

5.2 Comparison between the DDA models

The performance data of the second battery was tested with the non-parametric
Friedman test, which does not assume the independence of observations, as every
volunteer played for all DDA models applied during gameplay and the data is
 Title Suppressed Due to Excessive Length 13

not normally distributed. The analysis were made for each case, comparing the
four models for easy (1-4), normal (5-8) and hard (9-12) level sets. The data was
verified with the Shapiro-Wilk test, having significant p-values, failing to reject
the null hypothesis and concluding that data is not drawn from normal distribu-
tion for all the analyzed data. Moreover, it was observed with the Bartlett’s test
that samples from populations have equal variances for most of the questionnaire
data and not for performance data.
As seen in Figure 6, players numerically described the difficulty of levels for
three sets of four levels. From this, it was possible to observe that the mean
value grows according to the progression of the easy and medium levels, which
is expected in view of the difficulty to be increasing in platform games and the
algorithm for adjusting the levels to follow this pattern. However, there is a
peak at levels 5 and 9 compared to the previous growth pattern, as there is no
application of the DDA in them.

Fig. 6. Mean value and standard deviation of the difficulties measured by the ques-
tionnaire at each level of the second battery, divided into the three difficulty sets tested
and grouped by the four models.

According to the progression of the game, more data is captured and a better
estimate of the player’s performance and affective state occurs. Thus, in the
normal and hard level sets, a greater standard deviation is observed when the
application of the DDA does not occur. In addition, the difficulty of these levels
stands out in relation to the immediately following (6 and 10) because they fit
the player’s performance during the game until then, including the level without
DDA, and thus decrease the difficulty for the players average.
Additionaly to the questionnaires, 7 of the 12 levels were completed by all
volunteers. Besides, the levels with lowest percentage of completion of each set
were those without the application of DDA (levels 5 and 9 with 94.4% and
63.9%, respectivelly). Also, there were peaks in the affective and hybrid levels,
emphasizing that all participants completed the levels with affective adjustment.
14 M. Rosa et al.

The lowest percentage of completion with DDA was in the performance-based

level 10 (80.5%), which is still much higher than the lowest percentage without
DDA in level 9 (63.9%).
In addition to the increasing rate of players completing the levels, considering
performance data, a wider range of models medians difference were statistically
significant, only not for death rate and jump count on easy levels, considering a
significance level of 0.05. Table 1 presents a clipping of the cases for measured
performance data p-value with the Friedman test. Furthermore, the most consis-
tent significantly different performance data for different levels was playing time
by the participant on each level, shown in Figure 7. In this sense, statistics of
the players show a better performance with DDA models, specially considering
a lower time to complete the game and with less dispersion.

Difficulty Set Performance data Friedman

Easy playing time 0.0005
Normal Death rate 1.438e-10
Normal Jump fail count 2.5361e-06
Normal playing time 1.3800e-05
Hard Death rate 5.8626e-10
Hard Jump count 2.0277e-10
Hard Jump fail count 2.0782e-13
Hard playing time 1.5065e-13
Table 1. Friedman measured p-value, according to performance data to each set of
levels by difficulty.

Fig. 7. Box plot for playing time, respectively, for (a) easy, (b) normal, and (c) hard
levels, comparing for the four different models.

The standard deviation in games without DDA was considerably higher,

because they do not fit to each player and consequently bring a greater variation
in the performance of distinct players in such levels. In most cases, the mean
times in levels with different models of DDA are approximated in each analyzed
set. Besides, there is a negative peak in the standard deviation in performance
DDA in the easy and normal difficulty sets and in the hybrid DDA in the hard
set.
 Title Suppressed Due to Excessive Length 15

Moreover, at the end of each level, the player was questioned about which one
had a more adequate difficulty, shown in 8. As each level had a different model,
the models could be compared and had the following order of better perception
for the players: hybrid DDA, affective DDA, performance DDA, and without
DDA. It is observed that, for hard levels, it is easier to note a difference between
models because of greater disparities in player’s abilities than in easy and nor-
mal levels. However, the perception of the player to differentiate the difficulty
adequacy decreases in higher level sets, because of the increasing difficulty.

Fig. 8. Frequency of answers about the difficulty being more appropriate in the current
or previous level with frequencies grouped by difficulty sets, asked at the end of the
levels in the second battery.

Considering the data from questionnaires and from the performance data
presented above, the corresponding analysis can validate the presented hypothe-
sis. This is specially observed in the reduction of dispersion, higher performance
results and the shown perception of the player about the difficulty. Therefore,
a hybrid DDA system have a suitable difficulty adaptation for the player. Also,
it was perceived that future tests of comparison should be made separately to
be able to measure the difference between models through questionnaires and
maintain the player’s flow state.

5.3 Comparison between DDA distinct adaptations

Questions were made at the end of the third battery to verify if the player was
in a state of flow, and if the difficulty was adequate in their perception, as seen
respectively in Figures 9 and 10. In this sense, volunteers reported the combined
and platform adaptations as the best in average to maintain the player’s state
of flow, considering the focus on the game, losing track of time, and automatic
actions. The cases where no ADD was applied come next, with the adaptation
only on the jump having a lower rating.
The difficulty was found to be most appropriate for the platform adaptation,
followed by the combined adaptation and with lowest result for the game without
16 M. Rosa et al.

Fig. 9. Mean value and standard deviation of the questions related to the state of flow
of the third battery of tests, grouped by the four adaptation types.

Fig. 10. Mean value and standard deviation of the questions related to the adequacy
of difficulty of the third battery of tests, grouped by the four adaptation types.

DDA. Also, for the quality of the experience perceived by users when playing,
the DDA applied only to platforms had a better performance compared to other
models. It is also noticed that playing without any DDA application still yields
a better experience than the DDA applied to the jump or combined.
However, the data was also tested with the non-parametric Kruskal-Wallis
test, which assumes the independence of observations, as the data is not normally
distributed and it was guaranteed that each volunteer played only with one DDA
adaptation. For a significance level of 0.05, the differences between the medians
are not statistically significant for the majority of the questionnaire data.
As game levels are designed to grow with player performance, levels with
higher extremes are needed to actually analyze a difference when models are
applied in relation to player perceptions. In this sense, there must have been a
co-variance between the levels and the ADD model, with the levels varying in a
fixed way and ADD varying on top of that. To avoid this, future tests will have
levels revised to provide greater variability of adaptations.
In parallel, the players’ perception regarding the jumping mechanics in the
game was analyzed, as seen in Figure 11. The combined DDA turned out to be
 Title Suppressed Due to Excessive Length 17

the one whose jumping would be the most difficult to control and the easiest
would be with the jump DDA. In terms of the responsiveness and satisfaction
generated by the jump, platform DDA outperforms and without ADD was the
worst case, indicating that adaptation had an effect on the perception of controls.

Fig. 11. Mean value and standard deviation of the questions related to the jump control
of the third battery of tests, grouped by the four adaptation types.

In this sense, as in the second battery, the most consistent divergence of data
for different levels was the playing time to complete each level, shown in Table 2.
The playing time in all levels had a lower average and standard deviation for the
adaptations, with lowest average for the jump adaptation and lowest standard
deviation for the platform adaptation. Therefore, the adaptations were able to
decrease the variation in the performance of distinct players.

playing Time
Adaptation Average Median Standard Deviation
Without DDA 1164.00 808.39 1110.87
Platform DDA 1038.46 884.63 517.84
Jump DDA 1012.95 830.73 689.65
Combined DDA 1080.32 924.24 660.62
Table 2. Average and standard deviation of the playing time to complete the game
for each adaptation in the third battery.

The DDA adaptations aim to lower the dispersion and the average playing
time or death rate by fall, which is related to chasms or death areas. The results
have shown better adequacy related to the player preference for easy games.
Thus, Tables 3 and 4 present the average and standard deviation values for
playing time and death rate by fall. When considering performance data for
player’s with preference for easy games, the differences for playing time and
death rate were statistically significant when using Kruskal-Wallis test with 0.5
18 M. Rosa et al.

of significance level, and comparing the game without DDA with each distinct
adaptation (p-value between 0.018 and 0.023).

playing time
Adaptation Average Median Standard Deviation
Without DDA 3359.44 2776.09 1900.83
Platform DDA 1471.35 1166.59 842.73
Jump DDA 1119.11 1212.23 236.83
Combined DDA 1230.49 1257.71 267.89
Table 3. Average and standard deviation of the playing time to complete the game
for each adaptation with player’s that prefer easy difficulty in the third battery.

playing time
Adaptation Average Median Standard Deviation
Without DDA 50 85 25.02
Platform DDA 30 68 21.98
Jump DDA 31 75 14.83
Combined DDA 33 73.5 10.15
Table 4. Average and standard deviation of the total deaths by falls for each adapta-
tion with player’s that prefer easy difficulty in the third battery.

The platform and composite adaptations generated a better experience re-

lated to the state of flow, but without significant difference. Therefore, it was
inconclusive if the combined DDA was able to improve the maintenance of the
player in a state of flow when using the game without DDA as basis. On the
other hand, the combined DDA system was able to adequately tailor the diffi-
culty level, decreasing the dispersion of performances. The result was positive
specially for players that prefer the easy difficulty, showing that the system may
be improved by adapting the game according to the player profile.

6 Conclusion
The DDA studies are growing in scope of themes and studies carried out, con-
sidering several variables such as the player’s performance or the affective state.
However, few studies in the literature combine both factors, categorizing a hybrid
model. Specifically on the platform genre, no research with this model was found.
Furthermore, in-depth investigations of the efficiency of sensor use are scarce,
especially when considering real-time adaptation, and of hybrid adaptation to
keep the player in a state of flow.
Therefore, this work aimed to investigate a hybrid DDA model in platform
games. More precisely, the focus was on adapting the difficulty to each player
 Title Suppressed Due to Excessive Length 19

and keeping it in a state of flow, while the game was modified based on the per-
formance data and the player’s affective state, specifically through the analysis
of the Electrodermal Activity (EDA) of the skin. In addition, the performance-
based model was divided into three possible adaptations: the size and position
of platforms and gaps, jump height and the combination of those.
From the analysis of the obtained results, the method of computation of
difficulty associated with the developed DDA was corroborated. Thus, it was
observed a relationship between the difficulty estimated by the algorithm and
experienced by the players participating in the test. Also, it was possible to
balance the game in relation to the components used. Thus, the DDA test levels
remained linear and without the application of conditions that could bias the
player’s measure of difficulty and affective state, including level planning that
did not bring challenge by puzzles or traps.
Based on the results, limitations were observed when comparing few partici-
pants between the analyzed models, and the models being only tested together.
Still, it was observed that a DDA system was able to successfully tailor the dif-
ficulty of a platform game to heterogeneous players, which was demonstrated by
the larger number of completed levels, and the lower average and smaller dis-
persion of the time to complete them. In particular, the difficulty level was more
adequate when using the hybrid model, making the experience more suitable for
each player.
Specifically about the performance-based DDA system, all of the adaptations
were tested independently and with a wider range of participants. Thereby, the
limitations of the previous test have been overcome, and it was possible to verify
the adequacy of difficulty in the adaptations, according to the players perfor-
mance. However, the improvement of the player’s state of flow was not verified,
when considering that the questionnaire data was not significantly different from
the game without DDA for the gameplay characteristics of the state of flow.
Further studies will apply levels with extremes to augment the possibilities
of distinct adaptations and players perceptions. Moreover, fields to increase the
efficiency of the DDA model are the development of alternative EDA computa-
tion through both normalization and peak recognition, and the recognition of
the style of play by means of machine learning. Finally, one can analyze the
difficulty as influenced by the game design and not only by the presence and
characteristics of the components. In this way, the use of traps and puzzles
could be investigated, as well as the presence of specific missions and contents
to challenge the player.

References
1. Andrade, G., Ramalho, G., Gomes, A.S., Corruble, V.: Dynamic game balancing:
an evaluation of user satisfaction. American Association for Artificial Intelligence
(2006)
2. Aponte, M.V., Levieux, G.: Difficulty in video games: An experimental validation
of a formal definition. Proceedings of the 8th International Conference on Advances
in Computer Entertainment Technology (2011)
20 M. Rosa et al.

3. de Araujo, B.B.P.L., Feijó, B.: Evaluating dynamic difficulty adaptivity in shoot’em

up games. XII Brazilian Symposium on Computer Games and Digital Entertain-
ment (10 2013)
4. Blom, P., Bakkes, S., Spronck, P.: Modeling and adjusting in-game difficulty based
on facial expression analysis. Entertainment Computing 31, 100307 (07 2019)
5. Bontchev, B.: Adaptation in affective video games: A literature review. Cybernetics
and Information Technologies 16(3), 3–34 (2016)
6. Chanel, G., Lopes, P.: User evaluation of affective dynamic difficulty adjustment
based on physiological deep learning. In: Augmented Cognition. Theoretical and
Technological Approaches (04 2020)
7. Chang, D.M.J.: Dynamic difficulty adjustment in computer games. Proceedings of
the 11th Annual Interactive Multimedia Systems Conference (2013)
8. Chen, J.: Flow in games (and everything else). ACM magazine 50(4) (2007)
9. Compton, K., Mateas, M.: Procedural level design for platform games. American
Association for Artificial Intelligence (2006)
10. Csikszentmihalyi, M.: Flow: The Psychology of Optimal Experience. New York:
Harper & Row (1992)
11. Denisova, A., Cairns, P.: Adaptation in digital games: The effect of challenge ad-
justment on player performance and experience. Proceedings of the 2015 Annual
Symposium on Computer-Human Interaction in Play pp. 97–101 (2015)
12. Hintze, A., Olson, R., Lehman, J.: Orthogonally evolved ai to improve difficulty
adjustment in video games. Applications of Evolutionary Computation pp. 525–540
(2016)
13. Imre, D.: Real-time Analysis of Skin Conductance for Affective Dynamic Difficulty
Adjustment in Video Games. Doctoral thesis, Algoma University (2016)
14. Koens, E.: Generating non-monotone 2D platform levels and predicting difficulty.
Master’s dissertation, Utrecht University (2015)
15. Liu, C., Agrawal, P., Sarkar, N., Chen, S.: Dynamic difficulty adjustment in com-
puter games through real-time anxiety-based affective feedback. International Jour-
nal of Human-Computer Interaction pp. 506–529 (2009)
16. Lopes, R., Bidarra, R.: Adaptivity challenges in games and simulations: A survey.
IEEE Transactions on Computational Intelligence and AI in Games pp. 85–99
(2011)
17. Novak, J.: Desenvolvimento de Games. Cengage Learning (2011)
18. Schell, J.: The Art of Game Design: A Book of Lenses. AK Peters, 2 edn. (2014)
19. da Silva Valentim Mourato, F.J.: Enhancing Automatic Level Generation for Plat-
form Videogames. Doctoral thesis, Universidade Nova de Lisboa (2015)
20. da Silva Valentim Mourato, F.J., Birra, F., dos Santos, M.P.: Difficulty in action
based challenges: Success prediction, players’ strategies and profiling. Proceedings
of the 11th Conference on Advances in Computer Entertainment Technology pp.
1–10 (2014)
21. da Silva Valentim Mourato, F.J., dos Santos, M.P.: Measuring difficulty in platform
videogames. 4ª Conferência Nacional em Interação Pessoa-Máquina (2010)
22. Soares, R.T.: Biofeedback sensors in game telemetry research. SBC – Proceedings
of SBGames 2016 (2016)
23. Vollmers, C.G.K.: Dynamic Difficulty Adjustment In Games Using Physiology.
Master’s dissertation, Aalborg University Copenhagen (2018)