Context-driven Near-term Intention Recognition

Abstract

Recognizing the intention of others in real time is a critical aspect of many human tasks. This article describes a technique for interpreting the near-term intention of an agent performing a task in real time by inferring the behavioral context of the observed agent. Equally significant, the knowledge used in this approach can be captured semi-automatically through observation of an agent performing tasks on a simulator in the context to be recognized. A hierarchical, template-based reasoning technique is used as the basis for intention recognition, where there is a one-to-one correspondence between templates and behavioral contexts or sub-contexts. In this approach, the total weight associated with each template is critical to the correct selection of a template that identifies the agent's current intention. A template's total weight is based on the contributions of individual weighted attributes describing the agent's state and its surrounding environment. The investigation described develops and implements a novel means of learning these weight assignments by observing actual human performance. It accomplishes this using back-propagation neural networks and fuzzy sets. This permits early discrimination between different pre-categorized behavioral contexts/sub-contexts on the human-controlled agent such as a military or passenger vehicle. We describe an application of this concept and the experimentation to determine the viability of this approach.