Semantic reliability of multi-agent intelligent systems

Generally the concept of reliability has been interpreted as applied to hardware and software and has been based upon the assumption that a system can be decomposed into subsystems or components to which success or failure probabilities can be assigned assuming perfect semantic transactions between them, i.e., with no consideration to variation in the interpretation of the meanings of messages between various components. In multi-agent intelligent systems, where the agents interact with each other in capacities other than merely sending and receiving messages, cooperative decisions are made based upon beliefs, desires, intentions, and the autonomy of individual agents. In such cases, even if the components as well as the interconnections are error-free in the classical sense, there can be serious failures due to semantic variability and consequently the concept of reliability needs to be extended to semantics as well. This paper attempts to establish this new concept of semantic reliability and explore its relationship to the system reliability and information extraction processes. Here we examine the communication between agents and semantic error modes in multi-agent systems using Rao and Georgeff's belief-desire-intention (BDI) model of intelligent agents to decompose the semantic variation into its contributing parts from various subsystems comprising the agents. From this, the impact and the risk management strategies including fault tolerance are evolved. World representation, domain ontologies, and knowledge representation are brought out as important determinants of error control. A fault tolerance design based on goal hierarchy is suggested. (c) 2006 Lucent Technologies Inc.