The Impact of Humanoid Affect Expression on Human Behavior in a Game-Theoretic Setting∗ Aaron M. Rothy, Umang Bhatty, Tamara Aminy, Afsaneh Doryab, Fei Fang, Manuela Veloso Carnegie Mellon University, Pittsburgh, USA faaronr1, usb, tta, adoryab, [email protected], [email protected] Abstract motivated, we consider a competitive setting where a human will play a Stackelberg security game against a robot. With the rapid development of robot and other in- We aim to answer two research questions. First, in what telligent and autonomous agents, how a human way can a robot express a certain mood that can influence a could be influenced by a robot’s expressed mood human’s decision making behavioral model? To answer this when making decisions becomes a crucial question question, we create an NLP model to generate sentences that in human-robot interaction. In this pilot study, we adhere to a specific affective expression profile in the setting investigate (1) in what way a robot can express a of playing a competitive game. We equip a humanoid robot certain mood to influence a human’s decision mak- with verbal feedback system including these sentences. ing behavioral model; (2) how and to what extent Second, how does the observed affect expressed by a hu- the human will be influenced in a game theoretic manoid opponent impact a human being’s rationality and setting. More specifically, we create an NLP model strategy in a game theoretic setting? In behavioral game the- to generate sentences that adhere to a specific af- ory, an influential model of human’s behavior is the quantal fective expression profile. We use these sentences response model [GOEREE et al., 2016], and the parameter for a humanoid robot as it plays a Stackelberg se- in the model can be interpreted as the level of rationality of curity game against a human. We investigate the human. We run human subject experiments in a pilot study behavioral model of the human player. where we ask the human participants to play a Stackelberg security game [Yang et al., 2011] multiple times against the affect-expressive humanoid robot. We use Maximum Likeli- 1 Introduction hood Estimation to find the best value of the parameter that In the future, robots will be ever-present in our daily lives. We fits the data collected from human actions in the game plays, will live with them, work with them, and engage in all kinds and thus evaluate how the human behavior is impacted. of collaborative and potentially competitive interactions with them. Examples of robot sharing workspace with human can 2 Related Work already be found today in the realms of elderly care, rehabil- itation and healthcare, education, personal companions, and 2.1 Affective Expression in Humans and Robots social robots. In order to ensure that robot interact with hu- Emotions are an important aspect of human interaction. Hu- man in an intended way, it is crucial to gain a better under- man beings interpret emotion through nonverbal and verbal standing of how a robot affects humans behavior, especially cues. In [Frijda, 2005], Fridja argues that experiencing emo- in their decision making process. tions are, in fact, the primary purpose of social interactions. Humans evolved to read cues in emotions and moods. The Notably, it is well documented that the observation of another expressions of those in one’s surroundings can affect one’s person’s mood can have specific effects on behavior and can own levels of rationality, risk-taking, and decision-making. influence the observer’s mood following the interaction [Wild To what degree does this hold true for a robot companion as et al., 2001]. well? After all, a humanoid (or even non-humanoid) machine Affect is a general term relating to emotions, moods, feel- can act in a manner that humans can perceive as having an ings and other such states. Affective states vary in their de- “emotion” or personality. Social robots, in particular, can take gree of activation and valence (whether they are positive and advantage of uniquely human modes of interaction. negative)[Kirby et al., 2010]. In the psychological and cog- In this work, we seek to understand how a robot’s affect nitive science literature this is often represented via axes in expression can influence a human’s decision making. Thus a continuous multi-dimensional space [Russell, 2003]. Emo- tion is classified by short term, intense affective states and ∗This work was presented at 1st Workshop on Humanizing AI mood is classified by long term, diffuse states. Emotions are (HAI) at IJCAI’18 in Stockholm, Sweden. evaluative responses in specific events or to stimuli of impor- yThese authors contributed equally tance. Modeling emotion in a robot generally requires the interruption of an interaction, whereas modeling a mood al- on learning games such (e.g. memory and imitation). Cur- lows for the interaction to persist with only slight differences rently, mood contagion is an area that is receiving a great deal in the execution of the interaction. For these reasons most of interest, and although there still exists research on perfor- studies, including our own, focus on mood and not emotion. mance, it focuses on outcomes and is largely based on virtual People exchange verbal messages which contain informa- agents. To date, there have been no studies, that we know tion conveying their mental and emotional states. This in- of, which explore the influence of robotic affect on a player’s cludes the use of emotionally colored words and swear words. strategy, risk-taking, or rationality. Given the importance of affect in language, there has been a fairly substantial amount of research in affective statistical 2.2 Quantal Response language modeling[Wagner et al., 2014]. This includes the Quantal Response is a technique for extracting human ratio- development of affective NLG for generating medical texts nality [GOEREE et al., 2016]. We want to understand the [Mahamood and Reiter, 2011], and rule based emotive text error in an individual’s response and understand the probabil- generation based on sentence patterns [Keshtkar and Inkpen, ity distribution of the possible responses. If there are multi- 2011]. Notably, there has emerged work in the extension of ple actions in a situation, and the utility of a given action i is the LSTM (Long Short Term Model) language model for gen- Ui, then the best quantal response qi is calculated as follows, erating affective text [Ghosh et al., 2017]. where λ is used to indicate the deviation from optimality: Research has shown that human beings are capable of per- eλUi ceiving robotic affect under some circumstances. Different qi = n (1) P λUj forms of affective expression have been modeled with hu- j=1 e manoid robots. Bodily expression has been used for RO- We extend the work of [Yang et al., 2011] wherein the best MAN [Hirth et al., 2011], NAO [Hirth et al., 2011; Haring¨ response quantal response is used to figure out how rational a et al., 2011; Beck et al., 2012], KOBIAN [Zecca et al., 2009; player is in a Stackelberg Security game. We hold the defend- Pelachaud, 2009a], and Max [Pelachaud, 2009b]. These stud- ing strategy constant but aim to model the λ of each human ies demonstrated that people are generally capable of recog- player of our game. In this formulation, U in equation 1 rep- nizing affective states. Work has also been done to develop i resents the utility of selecting the ith gate in a given round. a parameterized behavior model in which behavior parame- Using maximum likelihood estimation, we can find the most ters controlled the spatial and temporal extent of a behavior likely value of λ for a given player given the utilities in the for mood expression. This includes models that enable the given round and the gate choices of the player over the all continuous display of mood in an interactive game [Xu et al., rounds in the game. We can then compare these λ values over 2014]. Work of this nature with robots is less developed than different players to see which players are more rational. A λ similar work with software agents, but it is becoming more of 0 would indicate ‘irrational’ behavior (selecting actions in common. However, research focusing on the effect of af- a uniform random manner), whilst a higher λ would indicate fective language models in conjunction with human - social a higher human rationality (selecting more ‘optimal’ actions). robot interaction has been limited (discussed below). It has been shown that affective expression can have pos- itive effects on humans during an interaction. For example, 3 Our Approach studies with the robot Vikia have demonstrated the effective- We run human-subject experiments. Participants will be pre- ness of an emotionally expressive graphical face for encour- sented with and play the game shown in figure 1. We based aging interactions with a robot[Bruce et al., 2002]. Other ef- this on the gate game found in [Yang et al., 2011]. 1 A se- fects include the way of interacting with a robot with emotive ries of “gates” is presented which the human can “attack.” facial expressions [Gockley et al., 2006], the effectiveness of First, they play against a computer opponent. (This serves assistive tasks such as learning and motivation given vocal as a baseline.) Then, they will play “against” a humanoid emotion expression[Kessens et al., 2009], user behavior dur- robot (a Softbank Pepper robot). When they play against the ing support tasks [Robins et al., 2009], and user mood. Of humanoid, the robot will act in either an encouraging or dis- particularly interest to us are studies in the format of interac- couraging manner.