A Cross-cultural Study of Playing Simple Economic Games Online with Humans and Virtual Humans Elnaz Nouri, David Traum Institute for Creative Technologies, USC HCII 2013 - 25th July Las Vegas Abstract We ran simple online economic interactions between virtual human and people from two countries (India and US). We compare our results to: Previously reported findings of similar interactions run in the laboratory. Questions we try to answer Opponent’s effect: How similar or different do participants feel and act when playing a virtual human versus another person? Culture’s effect: How different are players from the United States from Players in India? Game’s effect: What impact does the type of game have on players’ decisions and values? In this talk Previous work: Cultural differences in decision-making games and values MARV model and MARV survey Social aspects of human-agent interaction On-line AMT Study and experiment details US and Indian participants Human vs. Virtual Humans Dictator or Ultimatum game Results Cross-cultural differences in game play and personal values Game effect Opponent effect Conclusion Previous Work In-person games have been for understanding people's economic decision making behavior. (Camerer, 2003) Ultimatum game(Guth, 1983), Dictator game(Bolton, 1998), Prisoner’s Dilemma (Rapoport, 1965) … An example: Ultimatum Game Two players can split a certain amount of money (Güth, 1982). Two turn game: 1. 2. Proposer: make an offer Responder: accept or reject the offer Expected Results: Offer the minimum amount possible Accept any offer greater than zero Cultural Variations in Ultimatum Game Observations: Proposers offer about 40 – 50$ on average. Responders reject offers of 20$ or less. (Camerer, 2003) Considerable variation of offers and acceptance rates across 4 cultures (Roth 1993; Camerer 2003) (Roth, 1993) OFFER-100 OFFER-90 Israel OFFER-80 0 OFFER-70 Israel US OFFER-60 US Japan OFFER-50 0.1 Japan Yugoslavia OFFER-40 0.2 Yugoslavia OFFER-0 0.3 OFFER-30 0.4 OFFER-20 0.5 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 OFFER-10 0.6 MARV Model (Nouri, 2011) MARV = Multi-Attribute Relational Values Goal: Internal computational model of decision making for agents that is sensitive to culture, and produces behavior consistent with observations of that culture Approach: applying multi-attribute decision-making model to calculate utility of decisions • assigning appropriate weights to each of the following attributes: VS08up Hofestede Survey Methods for setting the weights Questions: 1. have sufficient time for your personal or home life 2. Q2 have a boss (direct superior) you can respect Previous Methods for determining culture-specific weights on attributes: (IDV) 1. Intuitions based on Hofstede’s dimensional (PDI) model of culture (Nouri & Traum 3. get recognition for good performance (MAS) 4. have security of employment (IDV) 2. Machine learning from human behavior (Nouri et al CogSci 5. datahave pleasant people2012) to work with (MAS) 6. do work that is interesting (IDV) New method Directly ask participants much by they 7. how be consulted yourcare boss inabout decisionsthe involving your work (PDI) weights (Nouri and Traum GDN 2013) 8.by assigning weights from -5 to 5. (-5 live in a desirable area (MAS) Hofstede dimensions of cultural values have a job respected by your about family and means they don’t care about that value whereas 9. 5 shows they care significantly thefriends value.) (IDV) 100 93.7 91.0 10. have chances for promotion (MAS) 90 PDI: Power Distance (large80.6vs. small), 11. keeping time free for fun (IVR) 69.6 80 73.9 IDV: Individualism vs. Collectivism, 12. moderation: having few desires (IVR) 68.0 13. being generous to other people (MON) 70 MAS: Masculinity vs. Femininity, 62.0 14. modesty: looking small, not big (MON) 60 UAI: Uncertainty Avoidance (strong vs. 15. If there is something expensive you really want to 42.7 46.0 49.2 US not calculated scores what do 50 buy but you do have Hofstede's enough money, weak), 40.0 you do? (LTO)India calculated Hofstede's scores 40 LTO: Long- vs. Short-Term Orientation, 29.0 16. How often do you feel nervous or tense?(UAI) 30 24.6 17. Are you a happy person? (IVR) IVR: Indulgence vs. Restraint, and 18. Are you the same person at work (or at school if 20 MON: Monumentalism vs. Selfyou’re a student) and at home? (LTO) 10 19. … 0.0 CMVC 2011) Effacement. 0 PDI IDV MAS UAI LTO IVR MON MARV Survey Abbreviation Value Description Vself Getting a lot of points Vother The other player getting a lot of points Vcompete Getting more points than the other player Vfairness having the same number of points as the other player Vjoint Making sure that if we add our points together we got as many points as possible Vrawls The player with fewest points (whoever that is) gets as many as possible Vlower bound Making sure to get some points (even if not as many as possible) Vchance The chance to get a lot of points (even if there's also a chance not to get any points) Rating scale: from -5 (no importance at all) to 5 (significantly imporant) Experiment Set Up Played online version Ultimatum Game or Dictator Game over 100 points The Ultimatum Game. as described. The Dictator Game. played exactly like the standard Ultimatum Game, except that the responder is not given an opportunity to accept or reject the offer. Paid based on their performance in the game: $0.5 show up fee Could earn another $0.05 for each additional 10 points that they accumulated in the game. Snapshot of the SimCoach character http://labs.simcoach.org/simcoach/?space=mini&character=3072 Experiment Procedure Before Game: 1. 2. Fill out the VS08 Hofstede Survey and demographic information questions Receive instructions about the game (Dictator Game or Ultimatum Game) denoting they would be playing with another participant from their country. The Game: Play the proposer in Dictator Game or Ultimatum game 2. Fill out the MARV Decision-making values survey 3. (in the case of the ultimatum game) Receive their partner’s move and their final reward. 1. Study Participants Indian and US participants recruited on Amazon Mechanical Turk. Number of Dictator Game players from India Ultimatum Game Human 107 101 Virtual Human 38 47 Number of players from US Dictator Game Ultimatum Game Human 107 101 Virtual Human 46 53 Results Results: Game Effect Offer Distribution Offers: The offers made in the two games are significantly different from one another. Average Offer in Dictator Game Average Offer in Ultimatum Game 39.6 47.6 • Values: are significantly different between the two games: {Vother , Vcompete, Vequal , Vjoint, Vrawls, Vlower bound} Results: Culture’s Effect Average US offers US Offers: 44.16 Average India offers India 41.44 Significant difference between the two cultures when playing ultimatum game with Virtual humans (p value<0.05). Values: Significance difference between the values reported by Indians and Americans (across all conditions): { Vself, Vcompete, Vchance} When playing with Virtual Humans { Vother in dictator game, Vlowerbound in ultimatum game} Results: Opponent’s Effect Offers: Playing against a virtual human or a human does not bring about significant difference in the offers made in the games. Indians played differently when playing ultimatum game with a virtual human as opposed to a human (p value<0.05). Values: Significant differences in the values reported { Vself, Vother, Vcompete, Vrawls, Vlower bound, Vchance} Prediction of offers Accuracy of prediction Percent correct Dictator Game Ultimatum Game Country (US or India) Hofstede Scores (7 dimensions) Hofstede Questions (28 questions) Decision-making Values (8 values) Random baseline: frequency of offers in the data 39.55% 39.56% 39.24% 52.86% 21% 51.61% 50.05% 53.52% 54.90% 32% Most common offer baseline (50%) 38% 51% Prediction of Culture percent correct Dictator Game Ultimatum Game Offers (11 values) Hofstede Scores (7 dimensions) Hofstede Questions (28 questions) Decision-making Values (8 values) 53.40% 64.42% 76.39% 60.09% 54.11% 69.85% 77.79% 65.70% Random Baseline 50% 50% Conclusion *** Our results are consistent with reported results in the literature. Opponent’s effect: People from US and India both treat virtual humans similar to how they would have treated another human. We conclude that virtual humans can be a reasonable substitute to humans in online economic interactions. (eg. Selling and negotiation) Culture’s effect: Values held by people from the two countries are different under similar conditions and the reasons should be further investigated. Game’s effect: The most prominent cause affecting the game behavior and the offer values is the type of the game being played. Future Work More data collection More cultures More types of games (potential for other values to be distinguished) Modeling Culture-specific agent models based on reported values Correlations between Hoftstede questions/dimensions and values (for cultures with no values data reported) Thank you! Questions?