MRL = Multirobot Learning

A quote from Multiagent Learning in Large Anonymous Games (Kash, Ian A et al.)

“With more learners, the noise introduced into pay- offs by exploration and mistakes becomes more consistent. Second, having more information typically improves performance. Publicly available statistics about the observed behavior of agents can allow an agent to learn effectively while making fewer local observations.”

So, for auctioning task would it be wise to be continuously auctioning off tasks to the robots even though they won’t actually be doing them so that they can learn how to bid more accurately in the different situations?  Use Dyna architecture.