Expressing and Understanding Desires in Language Games

摘要

We speak because we want to get certain things accomplished. In this study we present the simulation of a multia-gent language game which takes this into account. In contrast to previous language game studies, our agents use reinforcement learning to learn a function assigning a value to every state of the game. This value, that tells the agent how desirable the state is, is used along with a forward model to select actions. The agent can select verbal and non-verbal actions, depending on whether speaking or manipulating the world directly is more likely to bring about the change which the agent desires. On top of these capabilites, we used two rule-based agents to train a language learner. The learner trains a forward model of context-dependent utterance effects, which he then uses to express his desires and understand the desires of other players.