DeEvA, a Depot of Evolving Avatars

Abstract

This paper introduces the DeEvA platform for the generation of virtual characters. The platform generates virtual characters using personality traits as input. The generation process is conceived to generate characters whose physical appearance responds to people's expectations. Characters generated using the platform can be used as believable Embodied Conversational Agents in interactive applications. The platform uses a combination of crowdsourcing techniques: Reverse Correlation and Interactive Genetic Algorithms. This paper describes the method as well as three working examples.

Mapping symbolic descriptors to appearance

The method underlying DeEvA has been already described in detail in a previous publication [ 7 ] and it is summarised in the following.

The method is based on the use of Interactive Genetic Algorithms (IGAs) [ 3 ]; interactive because the computation of the tness function is based on the contribution of human beings at each iteration of the algorithm. The administrator of DeEvA con gures an experiment to nd correlations between a set of personality traits and a set of physical attributes. At the beginning, DeEvA generates a set of virtual characters by associating random values to the physical attributes. The characters are shown to human users, who vote using Likert scales (see Figure 1 for an example). The votes are used to compute the tness value of each character. When the users provide a su cient number of votes, DeEvA elects the best rated individuals for a migration to the next generation of the genetic algorithm. Additional individuals are generated by cross-over and mutation of the elected ones. The way users vote the individuals is a Reverse Correlation technique, which is used in the eld of psychology to nd correlations between traits and appearance. The combination of the Reverse Correlation with a Genetic Algorithm is conceived to improve the correlation at each iteration. The experiments published in [ 7 ] partially con rm the e ectiveness of this approach. DeEvA o ers also the possibility to \generate" virtual characters from a personality pro le. A pro le is a list of values, one for each trait, based on the same Likert scale used in the voting phase. The generation of a character is in fact implemented by searching for the individuals with the best tness values within the space of the selected pro le. Figure 2 shows an example. 3

Running experiments

This paragraph summarizes the results of two experiments, already presented in [ 7 ], and describes a new experiment recently conducted.

The experiment Trustworthiness and Dominance aimed at nding correlations between the aforementioned traits and two physical attributes: Age and Gender. The data analysis shows the potentiality of DeEvA in improving the correlation results through genetic evolution. The collected data present a correlation factor between Dominance and the combination Age+Gender of 0:56. The correlation factor increases to 0:61 when considering only the best 50% individuals, and raises to 0:77 when considering only the best 25% individuals. This result suggests that indeed the tness function selects the best representatives for a generation.

The experiment What a ects dominance? con rms the ability of DeEvA in supporting reverse correlation experiments. Coherently with previous research results, the data analysis shows that the perception of the dominance is correlated with the following physical attributes: the size of the bones of the chin, the inclination angle of the eyebrows, the rectangularity of the face, the width of the neck, and the width of the torso.

Our latest experiment, Does gender a ect Agreeableness? 4, aims at nding the correlation between six facets of the Agreeableness trait (trusting, straightforward, altruistic, compliant, modest, and kind hearted) with two physical attributes in uencing the perception of the gender (gender and breast size) and two other confounding variables (body proportion and age). Figure 1 shows a screenshot of the voting page. The age has been modulated in the range 0.40.6 (i.e., from 20 to 38 years old). The body proportions has been constrained to values which keeps the character to believable aesthetics (0.25 to 0.75). The gender and the breast size attributes are left to full range 0.0 to 1.0.

The experiment collected 68 votes from 14 di erent people, 6 males and 8 females, average 28.3 (standard deviation 8.0). Each user voted an average of 5.2 characters (standard deviation 9.2). The collected data have been analysed to discover how the physical attributes in uence the perception of the traits. The analysis consisted of tting a linear model predicting each trait separately. Table 1a reports the adjusted R-squared correlation value for the prediction of each trait using all the four physical attributes. In most cases, the Gender attribute has the biggest in uence, as shown in the last column, which reports the p-value for the Gender. Existing research already provided evidence of a positive correlation between the gender and real agreeableness of individuals [ 2 ]. This experiment suggests that, in judging from aesthetics, the perception of agreeableness is in uenced by the gender, at least in ve out of six of its facets.

In order to nd the prediction model which maximises the correlation factor between the physical attributes and each of the traits, a backward selection using adjusted R-squared has been conducted. Table 1b reports the best combination of predictors and what is the relative correlation factor. The results show that gender is responsible for the perception of most of the agreeableness facets, except for modesty. For some traits, there is a marginal improvement in the prediction when combining the gender with either breast size or age. 4 https://deeva.mmci.uni-saarland.de/individuals/vote/13 - November 12th, 2015

Conclusions

This paper presented a method for the generation of virtual characters from personality traits. The method uses a combination of two crowd sourcing techniques, reverse correlation and interactive genetic algorithms, to de ne a mapping between personality traits and physical attributes. In a previous work, the authors already demonstrated the potential of genetic algorithms to increase the performances of reverse correlation techniques. This paper presents the results of another experiment aiming at nding correlation between gender and the perception of agreeableness. Future work will focus on switching from a Likert rating to a pair-to-pair comparison in order to reduce voting biases. Future experiments will be conducted by taking advantage of massive crowdsourcing services in order to evolve over several generations.

The method has the potential of changing the production pipeline of lowcost character generation. By providing a symbolic description of a character's personality, the platform provides candidate avatars whose aspect ts with the context of application in a fraction of the time needed with traditional production practices.

References

1. Cassell , J.: Social Practice: Becoming Enculturated in Human-Computer Interaction . In: Universal Access in Human-Computer Interaction . Applications and Services, LNCS, vol. 5616 , pp. 303 { 313 ( 2009 )

2. Costa , Jr., P. , Terracciano , A. , McCrae , R.R. : Gender di erences in personality traits across cultures: Robust and surprising ndings . Journal of Personality and Social Psychology 81 ( 2 ), 322 { 331 ( 2001 )

3. Goldberg , D.E.: Genetic and evolutionary algorithms come of age . Communications of the ACM 37 ( 3 ), 113 { 119 ( 1994 )

4. Hassin , R. , Trope , Y. : Facing faces: Studies on the cognitive aspects of physiognomy . Journal of Personality and Social Psychology 78 ( 5 ), 837 { 852 ( 2000 )

5. Loyall , A.B. : Believable agents: building interactive personalities . Ph.D. thesis , Mitsubishi Electric Research Laboratories ( 1997 )

6. Mansour , S. , El-Said , M. , Rude-Parkins , C. , Nandigam , J.: The Interactive E ect of Avatar Visual Fidelity and Behavioral Fidelity in the Collaborative Virtual Reality Environment on the Perception of Social Interaction . In: Proceedings of the 10th International Conference on Communications . pp. 387 { 395 . WSEAS ( 2006 )

7. Nunnari , F. , Heloir , A. : Exploiting Reverse Correlation for the Generation of Virtual Characters from Personality Traits . INTETAIN 2015 , IEEE, Turin, Italy ( 2015 )