Efficient communication using reinforcement learning in a cooperative navigation game

Abstract

The thesis aims to investigate if agents are able to develop an efficient communication, with semantic meanings, and solve a navigation problem, using reinforcement learning. Additionally, it aims to evaluate the relevancy and benefit of one and two-way communication in comparison to each other and no communication. The problem is tackled in a multi-agent system (two agents), using a cooperative navigation game. The agents possess different privately held information, they are hence equipped with a communication channel and a language with no initial semantic meaning to convey the information to each other and solve the task of finding a target inside an environment with distracting obstacles. The experiments take place in both a discrete and a continuous setting with a varying number of communication ways and are evaluated based on the average time to complete the navigation. It is shown in the thesis that the agents can develop a language with a semantic meaning, which contributes to an efficient communication when set in a discrete environment and in a continuous static environment. However, it is inconclusive whether there are any significant benefits to be gained from a two-way communication compared to a one-way communication and whether the task can be solved in a continuous non-static environment.