Learning, Equilibria, Limitations, and Robots

Michael Bowling
CMU

The recent surge in research on multiagent learning has focused on
marrying ideas from reinforcement learning and game theory.  This
marriage has adopted the game theoretic concept of Nash equilibria as
one of its foundations.  Equilibria rely critically on the assumption
that all agents act optimally.  Reinforcement learning on the other
hand has tackled the scaling problem by sacrificing this very thing,
optimality, in favor of less memory and time.  This talk examines this
tension both from the game theory and reinforcement learning
perspectives.  I will briefly introduce the theoretical challenge that
limitations give to the existence of equilibria, and present
theoretical results that begin to address this challenge.  I will also
present results of combining my own WoLF variable learning rate
technique with learning techniques for scaling to large and complex
problems, such as an adversarial robot learning task.  These results
show that WoLF is a powerful tool for handling multiagent settings and
for scaling to large and realistic problems.