Cooperative learning in a multiagent society is characterized by the “what,” and not only the “how.” Agents would better learn how to satisfy goals set for them if they had better information. Experience is key: agents that have had more trials (and likely learned more) can teach others better. This is an assumption of the notion of weighted strategy sharing (WSS), which asserts that the more trials an agent experiences, the more expert it will become.

In this paper, WSS is intertwined with Q-learning to form the basis for a framework for multi-criteria expertness-based cooperative Q-learning (MCE). Levels of expertness are associated with agents in a society, and they are used to optimize the information exchange in a world of homogeneous agents that learn and explore individually and then collaborate to achieve a goal. This approach, applied in the world of mazes and the hunter-prey problem, is capable of speeding up the learning process and maximizing cumulative rewards. The authors examine a variety of parameters that influence the processes, including the role of temperature (as a measure of the randomness of actions taken) and the frequency of sharing. The authors suggest that all experiences, not only the successful ones, could be important experiences for agents in the context of MCE. Incorrect knowledge is another direction for consideration in further research on the topic.

This paper is an example of the efficient marriage of multiple existing approaches in research, making it an interesting read for researchers of multiagent societies, regardless of their disciplinary provenance (mathematics, computing, or cognitive sciences).