Node Correlation Effects on Learning Dynamics in Networked Multiagent Reinforcement Learning

Systems of intelligent interacting agents demonstrate high complexity of learning process due to complexity of single agent learning combined with their communication. Agent interactions are aimed at enhancing speed, quality, and complexity characterictics, nevetheless, each interaction may worse single agent results as well as enhance them. Therefore, building effective communication patterns is of high interest for learning process of intelligent systems. As an applied task, we consider project execution dynamics, where single tasks are assigned to employees having several conflicting parameters, while an intelligent system consists of multiple intelligent agents, learned by reinforcement algorithms. Different patterns of interaction according to agent similarities are explored as a factor affecting learning process. The condition of two agents connection is there similarity value, greater than some determined threshold; similarity function is determined for five static and dynamics parameters, and their influence is regulated by the corresponding five multipliers. The experiment shows there are significant parameters, showing more effect of connection on learning dynamics. This can be seen via effect of parameters, regulating neighbours contribution.