CONSMAG: Knowledge Representation Based Upon Consensus for Multiple Agents

A multi-agent system designed to observe and remember patterns of co-occurrences in an open, dynamic system is described. The agents are organized according to function, with the low-level agents storing observed cooccurrences of key concepts or events in memory until being recycled for further service, thus simulating fixed short-term memory. The mid-level agents retain a short history (the length of this history is determined by the number of low-level agents available) of co-occurrence phenomena. The high-level executive agent computes three different, related associative networks as a graphical representation of the co-occurrences recorded by means of cumulative consensus of the mid-level agents and responds to queries. Thus the CONSMAG system effects a transformation from episodic memory to a class of associative networks representative of semantic memory. This learning paradigm supports cost estimates for learning any given associative network, including modification of one network into another, to simulate a novice becoming an expert.