Emergence of Proto-Communication using Action Primitives Symbolized in Recurrent Q-Learning Agents
Abstract: In this paper, we suppose the gesture theory that is one of the theories on the origin of
language, which tries to establish that the speech originated from gestures. Based on the theory, we assume that “actions” having some purposes can be used as “symbols” in the communication through a learning process. The purpose of this study is to clarify what abilities of agents are necessary to acquire usages of the actions as the symbols. To achieve the purpose, we adopt a collision avoidance game and recurrent Q-learning (RQL) agents as the game players. Our simulations showed that the RQL agents can emerge a proto-communication to avoid collision with opponent by using visual information that the opponent turns its own eye away from the agent through the learning process. Further, we found that the RQL agents also obtain a rotational behavior to avoid collision by expanding the maximum number of learning iteration.
Keywords: Collision Avoidance Game; Emergence of Proto-Communication; Neural Q-Learning
(NQL) Agents; Recurrent Q-Learning (RQL) Agents; Symbolization of Action Primitives