Vicarious Value Learning and Inference in Human-Human and Human-Robot Interaction

摘要

Among the biggest challenges for researchers of human-robot interaction is imbuing robots with lifelong learning capacities that allow efficient interactions between humans and robots. In order to address this challenge we are developing computational mechanisms for a humanoid robotic agent utilizing both system 1 and system 2-like cognitive processing capabilities. At the core of this processing is a Social Affective Appraisal model that allows for vicarious value learning and inference. Using a multi-dimensional reinforcement learning approach the robotic agent learns affective value-based functions (system 1). This learning can ground representations of affective relations (predicates) relevant to interacting agents. In this article we discuss the existing theoretical basis for developing our neural network model as a system 1-like process. We also discuss initial ideas for developing system 2-like top-down/generative affective (semantic relation-based) processing. The aim of the symbolic-connectionist architectural development is to promote autonomous capabilities in humanoid robots for interacting efficiently/intelligently (recombinant application of learned associations) with humans in changing and challenging environments.