When creating videogames, designers adjust the game characteristics to optimize the experience of the target players. This design process is usually manual, and the amount of insight that videogame designers have over the player is limited. Identifying the psychological pro le of a videogame user can lead to speci c adaptation of videogame aspects like narrative, length or challenge. Additionally, performing this identi cation automatically can both leverage the game experience and reduce the amount of work needed for customizing the players' experience. This research describes an emergent methodology to automatically create the personality pro le of a player, and a prototype implementation. Additionally, a pilot study has been run, with preliminary but promising results.
Nowadays, videogames include several adaptive features that react to the player actions and strategies, and this produces more adapted experiences in which users can craft unique and comfortable experiences. These adaptations are usually based on ad-hoc approaches focusing on the task at hand, and react to the player performance and in-game decisions. Designers are constrained by the accessible data that a speci c game can provide, usually restricted to variables related to gaming performance instead of what the player expects from the experience. This limits the designer's options when aiming at enabling a closer player-focused design.
Player-speci c game adaptation has been tested in several games, but the implementations mostly rely on di culty changes. This usually imply translating adaptability into adding or reducing features like the number of enemies, speed ? This work has been supported by the CANTOR project (PID2019-108927RB-I00) funded by the Spanish Ministry of Science and Innovation; and by the project FEI
INVITAR-IA (FEI-EU-17-23) of the University Complutense of Madrid.
Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). of events or time allowed until the user manages to complete the level. This approach can lead to successful gaming experiences. However, more complex games like roleplaying games, which have more complex dynamics, cannot rely on this simpli cation for a rich customized experience.
In particular, one of the most important aspects in roleplaying games is the player personality pro le. Roleplaying games have a very important component of decision making, and the di erent players can take di erent options according to both their role and their own personality traits. Adapting roleplaying videogames to the particularities of these di erent types of personality can leverage the players' experience.
Against this background, this paper sets out as an exploratory e ort to provide a method to automatically identify speci c aspects of the players personality. The research is based on the hypothesis that it is possible to carry out automatic personality pro ling in a non-intrusive way in a roleplaying videogame. In order to deploy the methodology and test the validity of the hypothesis, a study where the player pro le was automatically acquired was run. The research used the Big Five personality model, and the overall identi cation of the personality traits was inferred from the player interaction with non-player characters in a virtual environment. The results were compared against the results of a running a validated Big Five questionnaire with the subjects.
The results (detailed in section 4) show that, while the exploratory character of this work does not yet permit to draw strong conclusions, the initial hypothesis of being able to identify user personality in an automatic non-intrusive way through videogames is plausible. 2
While not exclusively related to psychological models, the idea of automated
user pro ling has already been explored in several videogames. For instance,
Wii Fit [
Regarding psychological aspects, the particularities, types and features of
personality have been studied from several perspectives. For instance, Raymond
Cattell proposed a theory in which personality was divided into 16 factors [
Openness to experience describes how creative, adventurous and curious the person is. Those with high scores are more curious and open to new experiences. Conscientiousness refers to the e ciency and organization of a person. A low score indicates greater spontaneity. Extraversion is connected to how outgoing and energetic a person is. The higher the score, the more enthusiastic and outgoing a person is. Agreeableness describes how friendly and empathetic a person is. Neuroticism describes how sensitive and nervous people are. A low score indicates that the person is emotionally stable.
Based on the Big Five model, there exist several Big Five Inventories (BFI)
with di erent amount of elements, but always with the same approach and
objective. One of the most widely used and currently accepted questionnaire is the
44-item BFI created by Oliver P. John and Sanjay Srivastava [
These models have been previously used in videogames. Chris Bateman and
Richard Boon pro led four types of videogame players [
In order to test the hypothesis presented in section 1, a simple prototype of a roleplaying game has been implemented. The objective is to have subjects playing freely and create a user personality pro le based on the initial stages of the gameplay and the decisions they make.
The videogame has been created with Minecraft [
Among others, the prototype used two mods: JourneyMap, a customizable mini-map so that any tester can nd their way through the level, and Custom NPCs, a tool designed to create RPGs based on answering dialogues and completing quests. 3.1
First Pilot Study A complex, multiple-branch story was written for deploying the system. The story was composed as a set of dialogs. The Non-Player Characters (NPCs) were assigned dialogue trees that follow a common narrative (composed by 4 short stories). The di erent branches were written to cover the Big Five categories, adding or subtracting points from them as the user chooses an option, their personality mirroring their answer to speci c conversational stimulus. These points have been assigned to each of the dialog interventions based on the research team criteria.
This rst prototype sought to observe the initial response to the interactive
experience. For the same reason the experiment was conducted in person and the
researcher had to observe the player while he or she played without interfering
with his decisions, and write down the options he chose for later analysis. 6
samples were obtained. Once the user had nished playing, a questionnaire with
the original Spanish version of the Big Five Inventory was handed over [
Openess Conscientiousness Extraversion Agreeablenes Neuroticism Max Min 14 -10 7 -10 21 -5 12 -12
Once the results from the questionnaire and the calculation of the game interactions are obtained, a comparison was made to relate both values and get the error rates for each factor. The results were obtained by applying the arithmetic mean of the absolute values of each subject's error. The results can be observed in the Table 2.
Openness Conscientiousness Extraversion Agreeableness Neuroticism Mean Std. Dev.
16.94 8.73 21.79 21.3 23.06 12.25 14.08 13.25
It can be seen that the category with the highest error is the Extraversion. Agreeableness is the category where the results are most related and the data is concentrated around the average. Following these rst results, it was hypothesized that the source of error for each category is due to the value ranges of the user's decision range are not equidistant. This is based on the fact that the best results obtained correspond to Agreeableness, this being the only factor that ful lls this equidistance. This means that the user start with 0 points and add or subtract with the chosen dialog option, where 12 points in the gameplay is equal to 100 in the questionnaire and -12 is 0. Section 3.2 describes the changes applied in the second prototype of the research. 3.2
Second Pilot Study Based on the results and conclusions of the rst iteration, a number of changes were applied to address the limitations. For this purpose a modi cation of the gameplay was made.
The dialogues interventions were revalued to o set category's maximums and minimums, as shown in Table 3. In addition, it was decided to include values to Neuroticism category. Additionally, only those branches of dialogues that the player accessed were taken into account for the total sum. Otherwise, if if all the possible lines were taken into account, the results would be normalized against non-valid data points.
Openness Conscientiousness Extraversion Agreeableness Neuroticism Max Min
In addition, and in order to increase the access to potential players, for this
second iteration, a telemetry system for automatic capture and storage of user
data was implemented. Screen sharing systems were used for this purpose. Steam
Remote [
In the second pilot study, a total of 42 experiments were carried out. Figure 2 shows the values obtained in this second pilot by each player for each one of the categories in the questionnaire and gameplay.
As it can be seen in Figure 2, the results yield a very high dispersion along the Y-axis in some factors such as Conscientiousness. However, other factors concentrate their results around smaller aspects such as Openness. For Conscientiousness, the dispersion cover almost all the possible range. While this is in line with the usual dispersion of the personality traits in the Big Five model, it contrast with the less dispersed factors. It is assumed that this is because of the characteristics of the videogame: most of the evaluation points have strong Conscientiousness di erences, and they de ne the type of game that the player chooses. As such, the stories are better suited for a diverse range. On the contrary, aspects less involved in the narrative decisions (like Openness ) tend to show a more condensed set of answers.
In parallel with the analysis of the responses, the accuracy of the model (i.e. the similarity between the results of the gameplay and the results of the Big Five questionnaire) is shown in Figure 3, which shows the mean squared error (MSE) of the Big Five dimensions. In all cases except for Conscientiousness in story 4, the MSE value is below 20%. This leads to support the idea that, while the accuracy of the player pro ling process is not perfect, the gameplay helps to approximate the Big Five values for players. The higher error that can be observed for Conscientiousness is assumed to be caused by the same characteristic discussed earlier, namely the wider range of captured values for Conscientiousness in the gameplay due to the narrative particularities of the environment.
Figure 3 shows relatively positive results. However, it is concluded that a larger volume of samples of each of the factors is needed for more statistical validation. It can be seen that Conscientiousness is the factor with the worst results due to its scarce presence in the experience as opposed to Openness as indicated in the previous table. The current state of the project is still in its early stages and the general applicability of the results is yet to be fully determined. One of the most noteworthy points is the decision of a prospective assignment of personality values for the dialogue options, that is, why it was decided to establish some previous values for their subsequent balancing. In the presented prototype, the values were set by the authors, but this implies a certain level of noise in the results. Even with this limitation, the results are positive, but further re nements of the system require an additional methodology of expert-based assignment of the personality values.
Based on the hypothesis that personality can be analyzed in a non-intrusive way through videogames, an underlying hypothesis can be extracted, that there is a relationship between a player personality outside the videogame and the way they play within it. To validate it, it is possible to refer to the data extracted during the second iteration where a clear relationship between the results can be seen, with an error attributable to this lack of psychological basis in the assignment.
One of the most direct uses of the proposed method is, in parallel with the assignment of question values by experts, the use of machine learning algorithms to, based on the results of the Big Five Inventory questionnaire, assign automatically computed values. The relevance of this approach was tried in this research. In particular, the consistency of the results were tested by carrying out a Pearson correlation study between the results from the gameplay and the results from the Big Five Inventory questionnaire. The results were 0:17, p 0:35. These obtained values lack statistical signi cance, but it is assumed that this is due to the relatively low amount of data for the correlation to be signi cant. The study was completed with Support Vector Machines which was trained with the options selected by the users in the gameplay session and their results in the Big Five questionnaire. SVM (precision = 0:15) along with Random Forest Regressors (R2 < 0), con rm that the statistical signi cance of the data set is still too low due to the reduced amount of data points.
In short, it has been concluded that, at least while having such a small data set, it is not yet possible to automate the identi cation of the personality or to study the relationship between the two sets of responses through machine learning. 6
The paper has reported on a emergent study on the possibility of automated player personality pro ling in RPG videogames. A plausible methodology for automatically identi cation user personality in a non-invasive way through an interactive experience has also been proposed The results have been reasonably positive (but yet inconclusive) and they show that there is a possibility of nding a methodology with which to identify, with a reasonable error, the player's personality where the use of dialogues and their interactions seem to be of great importance. Still, the exploratory nature of the research leaves several ways of improvement as part of the future work.
The most straightforward improvement to the proposed methodology would be to gather more data and study the e ect of applying machine learning techniques to compute the best t value, for all the dimensions in the Big Five model, to the questions in the story (as introduced in Section 5). In any case, regarding the in uence of the speci c questions in the story, it is clear that the assigned values in the dialogue interactions must be validated by an expert because the actual values have been assigned by the researchers and there has not been any expert involved in the process. Related to this, checking the robustness of the results and the methodology requires to increase the amount of obtained samples.
Further research includes the need to test this methodology against other videogames and thus be able to verify that the hypothesis applies to di erent genres. It is believed that this kind of approach is not of application in general, since it would be hard to apply dialogue-based interaction in a car-racing game, for instance. However, we believe that the approach is not necessarily restricted to classic, adventure based roleplaying games. For instance, decisions based dynamics not speci cally rendered as dialogues could be a source of information as well.
Villar et al.