Mini-tutorial: Essential concepts of multi-agent learning

Gerald Tesauro
IBM T. J. Watson Research Center

This talk will provide an introductory description of the
problems and open issues in multi-agent learning.  Basic
concepts in reinforcement learning and game theory are
covered, followed by a survey of a variety of approaches
currently being studied by researchers.  These approaches
are grouped roughly into two categories: "naive" approaches
such as fictitious play and evolutionary game theory, that
ignore non-stationary of other agents' play, and more
"sophisticated" approaches such as WoLF and "strategic
teaching," that implicitly or explicitly allow for such
non-stationarity.