Multi-word-Agent Autonomy Learning Based on Adaptive Immune Theories

Abstract

A multi-word-agent autonomy learning model as a new online learning model is presented based on clonal selection theory and idiotypic immune network theory, where words simulate B cells and antigens involved in adaptive immune responses to construct autonomic immune word-agents under the framework of autonomy oriented computing. Moreover, relations between words simulate specific relations between B cells and antigens, where attributes of words simulate receptors, and strengths of these relations are defined as affinities calculated by matching receptors. The environment simulates the region in which the word-agents interact with each other and share information. The goal of the model is to optimize these relations iteratively by increasing affinities with hyper-mutation of receptors of B cell word-agents. Experimental results on dependency parsing show that not only accuracies are improved continually, but also, compared with MSTParser, high accuracies can be gained with less learning samples.