Consistent Dynamic-group Emotions for Virtual Agents. Abstract The use of computational models of emotion in virtual agents enhances the realism of these agents in a variety of domains, including virtual reality training and entertainment computing. We consider these two domains as prototypical for Multi-emotional-Agent-Systems (MeASs), which are the focus of this paper. MeASs typically include groups of agents organised into clusters, for example a special-force unit. While each agent in such a group has its own emotional model, resulting in realistic individual emotional behaviour, the group as a whole can show unrealistic emotional behaviour. Currently there is no method to enforce emotional consistency of a cluster of agents while allowing agents to have individual emotions. Our approach introduces an emotional-state component that is a separate step in the computational model of emotion used by individual agents. The introduction of this emotional-state component enables multiple architectures for group emotions. We evaluate these architectures and conclude that several enable consistent integration of individual emotions and group emotions. We believe that our research enables agent- and scenario designers to benefit from the individual realism a computational model of emotion brings to virtual agents, without losing group consistency. Furthermore, by choosing one architecture versus another, designers can trade-off quality of the group emotion for computational performance. We have implemented one of the possible architecture in a MeAS simulation environment. We show, using this simulation environment, how agent in a group emotionally influence each other (arriving at a group-level panic state) and how a simple strategy of a group leader can influence the emotion of the agents in the group (effectively calming the group). Related Work/References [1]A. Braun, S. R. Musse, L. P. L. de Oliveira and B. E. J. Bodmann. "Modeling Individual Behaviours in Crowd Simulation". In CASA Computer Animation and Social Agents, pp , May 2003, New Jersey, USA. [2] B. Ulicny, D. Thalmann, Crowd simulation for interactive virtual environments and VR training systems, Proc. Eurographics Workshop on Animation and Simulation’01, pp , Springer- Verlag, [3] N. Netten. Towards Believable Virtual Characters Using A Computational Model Of Emotion. Master 's Thesis, LIACS, Leiden University, [4] D. DeGroot and J. Broekens. Using Negative Emotions to Impair Gameplay. BNAIC, Joost Broekens, Niels Netten, Doug DeGroot {broekens, cnetten, - LIACS, Leiden University, Netherlands AS 1 EMS 1 AS 2 EMS 2 BMS 2 EMS Group BMS 1 AS 1 Leader BMS 2 EMS Group BMS 1 AS 1 AS 2 BMS 2 EMS Group BMS 1 AS 1 Leader BMS 2 EMS Group BMS 1 AS 2 AS 1 Leader EMS 1 EMS 2 BMS 2 BMS 1 AS 1 Leader EMS 1 EMS 2 BMS 2 BMS 1 AS 2 AS 1 EMS 1 EMS 2 BMS 2 BMS 1 AS 2 Figure 1. Architectures 1 (top-left), 2 (bottom-left), 3a (top-right) and 3b (bottom- right). Subscripts are used to denote different agents. Figure 2. Architectures 4a (top) and 4b (bottom). Both based on appraisal grouping Figure 3. Architectures 5a (top-left) and 5b (top-right) and 5c (bottom-left). All three are based on emotional communication. AS 1 EMS 1 EMS 2 BMS 2 BMS 1 AS 2 Advantages of Dynamic Emotional State Architectures Easy integration of the concept of emotion into groups of agents. Easy simulation of emotional influences from different sources (both sources from within the agent and sources from other agents nearby). Facilitate the simulation of different emotional strategies. 1) Start: a group of ‘A’ type (normal) agents and one B type (leader) agent. 2) A few ‘A’ type agents set to the panic emotion start to affect the group. 3) Group totally in panic cased by emotional communication. 4) Agent B (leader) is mixed up in the group and tries to calm down the panicked A type agents by showing a very strong emotion (very positive, un-aroused and dominant). Architectures that are made possible by our approach range from high-performing to high-quality and trade-offs are possible between the two. Communication-based architectures (Figure 4, column 5a and 5b) that use our approach have high quality and little extra design considerations for group- emotions. Architectures based on group sharing of the emotional-state or appraisal system (Figure 4, column 3a,b and 4a) scale better. Future work includes the extension of a MeAS simulation environment [3] to test the different architectures. We believe that our research enables agent- and scenario designers to benefit from the realism a computational model of emotion brings to individual virtual agents, without losing group consistency. Also, designers can trade-off quality of the group emotion for computational performance. Figure 4. Comparison of architects (performance, quality and design effort) Overview of the Different Architectures The use of an emotional-state as emotion currency, made possible by separating the computational-model of emotion in three steps: appraisal, emotional-state maintenance and emotional behaviour (based on the same architecture as proposed in [4]), enables consistent emotions for dynamic groups of virtual agents. Here below you see different architectures that are possible when using an emotional state as emotion currency. Each architecture has some pro’s and con’s regarding the criteria as can be seen in Figure 4. AS 1 EMS 1 EMS 2 BMS 2 BMS 1 AS 2 Architecture Evaluation Test Implementation of Architecture 5c (see short demo) To test and experience the effects of an emotional communication architecture embedded within individual agents that form a group we implemented a test scenario with two types of agents (normal vs. leader) using architecture 5c. All agent influence each others emotion directly. Only leader type agents have a (cognitive) strategy to control the situation, i.e., keep calm or stay strong when confronted with a group of panicked agents.