Reasoning with Categories for Trusting Strangers: a Cognitive Architecture

Download

PDF - Accepted Version Download (647Kb)

Description/Abstract

A crucial issue for agents in open systems is the ability to filter out information sources in order to build an image of their counterparts, upon which a subjective evaluation of trust as a promoter of interactions can be assessed. While typical solutions discern relevant information sources by relying on previous experiences or reputational images, this work presents an alternative approach based on the cognitive ability to: (i) analyze heterogeneous information sources along different dimensions; (ii) ascribe qualities to unknown counterparts based on reasoning over abstract classes or categories; and, (iii) learn a series of emergent relationships between particular properties observable on other agents and their effective abilities to fulfill tasks. A computational architecture is presented allowing cognitive agents to dynamically assess trust based on a limited set of observable properties, namely explicitly readable signals (Manifesta) through which it is possible to infer hidden properties and capabilities (Krypta), which finally regulate agents' behavior in concrete work environments. Experimental evaluation discusses the effectiveness of trustor agents adopting different strategies to delegate tasks based on categorization.