Modeling enjoyment preference from physiological responses in a car racing game Simone Tognetti, Maurizio Garbarino, Andrea Bonarini, Matteo Matteucci Abstract— We propose a framework to estimate player enjoy- ment preference from physiological signals. This can produce objective measures that could be used to adapt dynamically a game to maintain the player in an optimal status of en- joyment. We present a case study on The Open Racing Car Simulator (TORCS) video game. In particular, we focus both on the experimental protocol, which we designed with special attention to produce physiological responses related to the game experience only, and on signal analysis, which produces a simple and general model good enough to estimate player enjoyment preference in real applications. I. I NTRODUCTION ”A good game is the one you like to play”. With this motto in her mind, the digital game designer conceives rules and structure of a game to maximize the level of enjoyment for the target audience. Realistic ambiance and characters, intelligent and reactive opponents with a human-like behavior can provide enjoyable game experience, however this cannot be assumed a priori. Some evaluation can be done by considering performance parameters, often assuming that a better performance is related to higher enjoyment of the game experience; but, different players may react differently to game features, and enjoyment has to be considered as a personal experience. According to the affective computing, we can assume that physiological response can be related to enjoyment [1], and that it can be taken as an objective measure of it. Thus, we present in this paper the application of a methodological framework for the estimate of preference among game experiences, from the physiological state of the player, in the car racing video game scenario implemented by TORCS – The Open Racing Car Simulator [2] . With the purpose of comparing and validating the pro- posed game scenario with respect to recent literature, we have carried out a correlation analysis between physiological data and subject preference during different variants of the game. Results show that features derived from some of the physiological signals (e.g., Galvanic Skin Response (GSR), Blood Volume Pulse (BVP) and Respiration (RESP)) have a high correlation with player reported preferences. On the other hand, as we expected from other studies in literature, only some physiological features show relevant high discriminating power. Therefore, signals such as Heart Rate (HR) or temperature are not really suitable as emotional input because of their poor correlation with user preference. Supported by these findings, we have estimated a linear Politecnico di Milano IIT Unit, Dip. di Elettronica e Informazione, via Ponzio 34/5, 20133 Milano, Italy. E-mail: {tognetti,garbarino,bonarini,matteucci}@elet.polimi.it. model based on physiological signals able to predict the subject enjoyment preference during the game. Our approach focuses on the differential comparison of preference between game situations and the resulting model can be used, in a future experiment, to modify at runtime the game experience accordingly to the predicted user preference to optimize user enjoyment. The ground truth about subject preference has been eval- uated by questionnaire analysis. The players are asked to express a preference between two variants of the video game. This approach of eliciting emotion is named comparative affect analysis and it was first introduced by Yannakakis and Hallam [3], [4]. Affective models are then derived using preference learning techniques [5], [6] matching user reported preferences and features from physiological sig- nals measured during the game session. We assume that physiological responses are not task dependent since the level of physical activity required for the interaction with our game is constant during the session. As presented by Yannakakis [7], the preference model can be learned using different computational methods. In this paper, we propose a different linear model obtained with Linear Discriminant Analysis (LDA) [8], which shows performance analogous to other models presented in literature [9], but lower complexity and lower computation demand. In the next sections, we first give an insight into the state of the art, then we introduce the experimental protocol designed for our experiments, finally we present the methods we adopted for preference modeling and the results obtained. A. State of the art With the purpose of determining criteria that contribute to player satisfaction, Yannakakis and Hallam [10] proposed two techniques for modeling player satisfaction in real- time. They assume that player-opponent interaction primarily contributes to the entertainment in a computer game. There- fore, metrics based on qualitative considerations of what is enjoyable from in-game performances (e.g., time before the player loses a life) have been considered as indicators of the level of interest. Another approach consists in modeling the entertainment by following the theoretical principles described by Mal- one [11], and concepts related to the Theory of Flow [12]. Qualitative factors such as challenge, fantasy and curiosity are the ones that, according to them, mostly account for player entertainment. Quantitative measures for challenge, curiosity and flow state can be derived from an empirical analysis of player responses to game mechanisms. All the 978-1-4244-6297-1/10/$26.00 c 2010 IEEE 321