Application to the Ultimatum Game Two turn game: Two players can split a certain amount of money. Proposer: make an offer (here from 0 to 100) Responder: accept or reject the offer Possible Outcomes: If accepted – money is split according to deal If rejected both get nothing A Cultural Decision Making Model for Virtual Agents Computer Science Department, USC Model’s Cultural Working Paradigm Modeling Culture Hofstede's proposed a dimensional model for culture based on the following dimensions: Masculinity (MAS), Individuality (IDV), Power Distance (PDI), Uncertainty Avoidance (UAI), Long Term Orientation (LTO) (Hofstede, 2001). We use cultural dimensional scores to contribute to the weight of each valuation . The weight of each valuation function is the product of the contirbution of each individual cultural dimen- sion. Our Model Our model has a multi attribute paradigm: and takes into account several metrics other than self interest: 1. Self Interest (the agent's own utility) 2. Other Interest (the utility of other parties involved) 3. Total Utility (sum of individual utilities of all parties) 4. Average Utility 5. Relative Utilities (viewed in several ways, such as self/total, self-other, self/other, self/average) 6. Minimum Utility (lower bound for any participant - Rawls' theory of justice) Motivation Most classical economic game-theory accounts of decision-making look at a monolithic notion of utility and maximizing expected utility as the key to rationality. However, it has been shown by social scientists that in practice game-theory models cannot predict how people make deci- sions and that people from different cultures behave differently in interactive situations (Camerer, 2003). Our Goals In order to build agents that can interact with humans we need to have a model that is good at replicating human behvaior. We want to build a decision-making model that can: a) replicate human behavior taking into account more than just Self-Utility b) capture cultural differences in the behavior of people from different backgrounds c) be incorporated into virtual agents Elnaz Nouri, David Traum Human playing ultimatum game with virtual agent Virtual agents playing ultimatum game Comparison of the model results with humans playing ultimatum game Results Summary We build a novel decision-making model that a) takes into account multiple valuations of a decision b) captures different ways of evaluating the same choice set across different circumstances c) takes into account individual and group differences in cultures d) is useful for modeling decision making in virtual humans interacting with each other and with real humans e) is explicit enough to implement and cover common situations Future Work a) Learn the weights using machine learning techniques, e.g. inverse reinforcement learning. b) Verify the hand-crafted rules. c) Incorporate into our model information about emotions, personality, and dialogue context. d) Focus on cultural differences with respect to persuasion. References Nouri, E., & Traum, D. (2011). A cultural decision-making model for virtual agents playing negotiation games. In Proceedings of the International Workshop on Culturally Motivated Virtual Characters. Reykjavik, Iceland Camerer, C. F. (2003). Behavioral game theory - Experiments in strategic interaction. Princeton University Press. Hofstede, G. H. (2001). Culture’s consequences: Comparing values, behaviors, institutions, and organizations across nations. Thousand Oaks, CA: SAGE. Acceptance Ratio Offer distribution Four Different Cultures Human - US Culture Cultural Background Affects Decision Making Rational Self Interested Agent Observed Behavior in Game Incorporating our Model into Virtual Agents General Table of Weights Example: Hand crafted MAS weights Setting the Model Weights Culture PDI IDV MAS UAI Virtual Human Mean Offer Human Mean Offer Virtual Human Rejecon Rate Human Rejecon Rate Hofstede's Dimensional Values Israel 13 54 47 81 $21.66 $41.71 25% 17.70% Japan 54 46 95 92 $32.50 $44.73 1% 19.30% Chile 63 23 28 86 $33.13 $34.00 1% 6.70% Austria 11 55 79 70 $33.13 $39.21 9.10% 16.10% Ecuador 78 8 63 67 $33.13 $34.50 1% 7.50% Germany 31 64 61 60 $36.88 $36.70 9.10% 9.50% US 40 91 62 46 $41.88 $42.25 12.00% 17.20% Spain 57 51 42 86 $28.75 $26.66 25.00% 29.20% Sweden 34 70 4 26 $33.13 $35.33 10.20% 18.20%