Scaling Up of Action Repertoire in Linguistic Cognitive Agents

摘要

We suggest the utilization of the Modeling Field Theory (MFT) to deal with the combinatorial complexity problem of language modeling in cognitive robotics. In new simulations we extend our previous MFT model of language to deal with the scaling up of the robotic agent's action repertoire. Simulations are divided into two stages. First agents learn to classify 112 different actions inspired by an alphabet system (the semaphore flag signaling system). In the second stage, agents also learn a lexical item to name each action. At this stage the agents will start to describe the action as a "word" comprised of three letters (consonant - vowel - consonant). The results of the simulations demonstrate that: (i) agents are able to acquire a complex set of actions by building sensorimotor concept-models; (ii) agents are able to learn a lexicon to describe these objects/actions through a process of cultural learning; and (iii) agents learn actions as basic gestures in order to generate composite actions.