| |
Abstract:
We are frequently called upon to do multiple tasks that
compete for our attention and resource. Often we know the optimal
solution to each task in isolation. How can we exploit that
knowledge to efficiently determine good solutions for doing the
tasks in parallel? We formulate this question as that of
dynamically merging multiple Markov decision processes (MDPs), and
present a new theoretically-sound dynamic programming algorithm
that assumes known good solutions (value functions) for the
individual MDPs in isolation and efficiently constructs a good
solution for doing the set of MDPs in parallel. Our algorithm can
merge MDPs dynamically, assimilating a new MDP smoothly into an
ongoing merging of previous MDPs. We analyze various aspects of our
algorithm and illustrate its use on a simple merging problem.
|
