Empirically Evaluating Multiagent Learning Algorithms
By Erik Zawadzki
An important part of real-world environments is that they are home to
multiple parties, each with their own goals and agendas. These other
agents make learning in these environments distinct from learning in
single agent settings: for instance, while your algorithm is learning,
it is simultaneously also teaching.
Research into multiagent learning has attracted a lot of focus recently.
This attention has created a wealth of new algorithms, but a lack of
general understanding of how these algorithms perform empirically. I
will describe an experiment, unprecedented in terms of scale, which will
cast some light on how these algorithms perform. Additionally, I will
talk about the unique analytical challenges posed by multiagent learning
algorithms and the methods that were used to analyze their performance.