Towards an emotion core based on a hidden Markov model

Abstract

An emotion core for autonomous robots based on a hidden Markov model is proposed. Different emotional robot characters can be designed by tuning state transition probabilities. Perception of stimuli has an impact on emotional state transitions, and, thus, affects emotion dynamics and observable expressions/actions. This paper proposes the methodology of design and implementation, and shows integration into a decision and control architecture. The application potential in emotion-based human-robot interaction is discussed.

Interacting with an artificial partner: modeling the role of emotional aspects

“…Principally, while our model has a strong timevarying imprint, those frameworks are mainly static. The dependence of transition probabilities on past inputs is considered in (Kuhnlenz and Buss 2004) through the implementation of a digital (forgetting) filter. However, modifications induced by inputs are not permanent: the impact of a stimulus is effective only during the time interval corresponding to the length of the filter response.…”

“…1. Our model differs also from that in (Kuhnlenz and Buss 2004) for other architectural details: our concepts of personality and attitude can be seen as unifying the two matrices of the HMM in a simpler structure; moreover, in our model emotional states are not hidden but directly output to the outside.…”

“…happiness_state) structures, upon which a set of behavioral responses is built (e.g. Kuhnlenz and Buss 2004). Being our work focused on modeling emotional interaction in a robotic context, rather than human emotions per se, in what follows we will use the terms emotion and emotional state without pretense to address the whole complexity of these phenomena.…”

“…In a PFSA the transition function is stochastic: that is, given the current state and the current input, there are many possible next states, each entered with a given probability (Rabin 1963;Paz 1971). Indeed, the introduction of a stochastic component in an emotional interaction model leaves space for unexpected behavior (Chittaro and Serra 2004;Kopecek 2003;Kuhnlenz and Buss 2004;Nomura 1996): we do not expect that our interlocutor will always react the same way to the same situation, even if her personality remains unchanged.…”

“…Another stochastic model, proposed as an emotional core for a robot, is a hidden Markov model (HMM), equipped in addition with input control (Kuhnlenz and Buss 2004). Here, emotional states are the HMM's hidden states, to which different observable expressions are associated with different probabilities.…”