Generating the Minimal Optimal Language for Cooperative Agents

Inspired by the advantages of languages used in human beings and other creatures, we attempt to create a novel formal language for autonomous agents in a cooperative setting. With a few assumptions, we define a minimal optimal language generation problem which is formulated as a complex optimization problem, where a set of words is automatically generated. These words are derived from symbols extracted from the agents’ perception of the environment. Sentences are formed from word combinations and attached to semantic meanings used to distinguish the situations when communication is required. We then, present an optimal algorithm and an approximate algorithm for generating the smallest word set. Search skills and heuristics are developed to optimize the language generation process. The performance of the proposed approaches is evaluated in several environments in simulation. Finally, we show how the minimal language can be used by agents during planning to coordinate their behaviors and conduct hundreds of task instances to illustrate the benefits of using the language in two domains.