Crucial Factors Affecting Cooperative Multirobot Learning

The effectiveness of multirobot learning in achieving optimal, cooperative solutions is potentially affected by various factors having to do with the nature and configuration of the robots and the nature of the learning entities. Varying one factor wrongly may lead to undesirable results.  There is no reported work on how systematically to set up these factors. In this paper, we methodically test the effect of varying four common factors (reward scope, global information delay, diversity of robots, and number of robots) in a decentralized multirobot system, first in simulation and then on real robots. The results show that two of these factors, reward scope and global information delay, if set up incorrectly, can prevent optimal, cooperative solutions.