Multi-Agent Communication and Behaviour Training using Reinforcement Learning. 01/11/2019 - 31/10/2023

Abstract

Many real-world applications require intelligent cooperative agents that can work together to solve a problem. An example of such an cooperative multi-agent application is the control of multiple autonomous vehicles. Multi-agent reinforcement learning is a wellresearched topic and many solutions exist in the state-of-the-art. Recently, the research community was able to create agents that learn how to communicate with each other to reach their goal. This is a new subfield of the multi-agent reinforcement learning domain in which we will research how we can achieve decentralised training of these communicating agents. This will allow us to create heterogeneous agents that can communicate or continue to train the communicating agents after their deployment which is not possible with the current state-of-the-art methods. In this project, I will extend the state-of-the-art by investigation how we can communicate with an unknown number of other agents which is a problem with state-ofthe- art methods. Next, I will work on the feedback structure that is used to train the communication between the agents. After the feedback structure I will work on splitting the agents architecture to create environment specific agents. I hypothesise that this will decrease the training time of the communication policy which is required for decentralised training. These advancements will be combined to create agents that we can train in a decentralised setting.

Researcher(s)

Research team(s)

Project type(s)

  • Research Project

Multi-agent communication and behaviour training using distributed reinforcement learning. 01/01/2019 - 31/10/2019

Abstract

Many real-world applications require intelligent cooperative agents that can work together to solve a problem. An example of such an cooperative multi-agent application is the control of multiple autonomous vehicles. Multi-agent reinforcement learning is a well-researched topic and many solutions exist in the state-of-the-art. Recently, the research community was able to create agents that learn how to communicate with each other to reach their goal. This is a new subfield of the multi-agent reinforcement learning domain in which we will research how we can achieve decentralised training of these communicating agents. This will allow us to create heterogeneous agents that can communicate or continue to train the communicating agents after their deployment which is not possible with the current state-of-the-art methods. In this project, I will extend the state-of-the-art by investigation how we can communicate with an unknown number of other agents which is a problem with state-of-the-art methods. Next, I will work on the feedback structure that is used to train the communication between the agents. After the feedback structure I will work on splitting the agents architecture to create environment specific agents. I hypothesise that this will decrease the training time of the communication policy which is required for decentralised training. These advancements will be combined to create agents that we can train in a decentralised setting.

Researcher(s)

Research team(s)

Project type(s)

  • Research Project