“Less work has been done on formal methods of decomposing tasks (and behaviors) into subtasks (sub-behaviors) appropriate for multi-agent solutions, how agents’ sub-behaviors interact, and how and when learning of these sub-behaviors may be parallelized.” (An Overview of Cooperative and Competitive Multiagent Learning)
I think that this would be an interesting problem to explore. I found that the problem of decomposition is also a major problem in engineering. They have come to the point were the number of variables in a problem are in the 1000s and an expert is not always available. Based on the paper, The State of Problem Decomposition in Engineering Design, solving this problem is useful. So, I am going to research how others are solving the decomposition problem and combine that with multiagent learning to try to solve real life engineering problems.