Simplifying the Design of Human-Like Behaviour: Emotions as Durative Dynamic State for Action Selection

Human intelligence requires decades of full-time training before it can be reliably utilized in modern economies. In contrast, AI agents must be made reliable but interesting in relatively short order. Realistic emotion representations are one way to ensure that even relatively simple specifications of agent behavior will be expressed with engaging variation, and those social and temporal contexts can be tracked and responded to appropriately. We describe a representation system for maintaining an interacting set of durative states to replicate emotional control. Our model, the Dynamic Emotion Representation (DER), integrates emotional responses and keeps track of emotion intensities changing over time. The developer can specify an interacting network of emotional states with appropriate onsets, sustains, and decays. The levels of these states can be used as input for action selection, including emotional expression. We present both a general representational framework and a specific instance of a DER network constructed for a virtual character. The character’s DER uses three types of emotional state as classified by duration timescales, keeping with current emotional theory. We demonstrate the system with a virtual actor. We also demonstrate how even a simplified version of this representation can improve goal arbitration in autonomous agents.