Culture-specific models of negotiation for virtual characters: multi-attribute decision-making based on culture-specific values

We posit that observed differences in negotiation performance across cultures can be explained by participants trying to optimize across multiple values, where the relative importance of values differs across cultures. We look at two ways for specifying weights on values for different cultures: one in which the weights of the model are hand-crafted, based on intuition interpreting Hofstede dimensions for the cultures, and one in which the weights of the model are learned from data using inverse reinforcement learning (IRL). We apply this model to the Ultimatum Game and integrate it into a virtual human dialog system. We show that weights learned from IRL surpass both a weak baseline with random weights and a strong baseline considering only one factor for maximizing gain in own wealth in accounting for the behavior of human players from four different cultures. We also show that the weights learned with our model for one culture outperform weights learned for other cultures when playing against opponents of the first culture.