Reinforcement Learning for Landmark-based Robot Navigation
----------------------------------------------------------

In previous work, we described a cooperative multi-agent system (MAS)
for coordinating various subcomponents of a robot navigation system.
In this system, manually-tuned bidding functions controlled the
interaction between different agents to guide a robot to a desired
target location.  A difficulty with hand-tuning is that it is hard to
handle situations involving complex tradeoffs.  This paper explores
the suitability of reinforcement learning for automatically tuning
agents within an MAS to optimize a complex tradeoff.  Based on
high-fidelity simulations, we conclude that reinforcement learning can
learn better bidding functions than our hand-coded procedures.  The
learned bidding functions yield higher probability of success, shorter
navigation trajectories, and reduced use of expensive sensing
actions.  However, we also found that correct functioning of
reinforcement learning required extensive experimentation with the
state representation.