Learning Strategies for Evolved Co-operating Multi-Agent Teams in Pursuit Domain
This study investigates how genetic programming (GP) can be effectively used in a multi-agent system to allow agents to learn to communicate. Using the predator-prey scenario and a co-operative learning strategy, communication protocols are compared as multiple predator agents learn the meaning of commands in order to achieve their common goal of first finding, and then tracking prey. This work is divided into three parts. The first part uses a simple GP language in the Pursuit Domain Development Kit (PDP) to investigate several communication protocols, and compares the predators' ability to find and track prey when the prey moves both linearly and randomly. The second part, again in the PDP environment, enhances the GP language and fitness measure in search of a better solution for when the prey moves randomly. The third part uses the Ms. Pac-Man Development Toolkit to test how the enhanced GP language performs in a game environment. The outcome of each part of this study reveals emergent behaviours in different forms of message sending patterns. The results from Part 1 reveal a general synchronization behaviour emerging from simple message passing among agents. Additionally, the results show a learned behaviour in the best result which resembles the behaviour of guards and reinforcements found in popular stealth video games. The outcomes from Part 2 reveal an emergent message sending pattern such that one agent is designated as the "sending" agent and the remaining agents are designated as "receiving" agents. Evolved agents in the Ms. Pac-Man simulator show an emergent sending pattern in which there is one agent that sends messages when it is in view of the prey. In addition, it is shown that evolved agents in both Part 2 and Part 3 are able to learn a language. For example, "sending" agents are able to make decisions about when and what type of command to send and "receiving" agents are able to associate the intended meaning to commands.