Markov Decision Processes (MDPs) allow us to build policies for maximizing the expectation of an objective in stochastic environments where the state of the world is fully observed. Partially Observable MDPs (POMDPs) give us machinery for planning when we have additional uncertainty over the state of the world. We can make a similar jump from MDPs to characterize uncertainty over other elements of the environment. Namely, we can also have uncertainty over the transition and reward functions in an MDP. Here, we introduce new classes of uncertain MDPs for dealing with these kinds of uncertainty and present several folk theorems showing that certain subsets of these can be reduced to the standard POMDP with uncertainty only over the state. We are particularly interested in developing these frameworks to explore applications to Negotiable Reinforcement Learning, a method for dynamically balancing the utilities of multiple actors.
Title
Uncertain Reward-Transition MDPs for Negotiable Reinforcement Learning
Published
2017-12-16
Full Collection Name
Electrical Engineering & Computer Sciences Technical Reports
Other Identifiers
EECS-2017-231
Type
Text
Extent
45 p
Archive
The Engineering Library
Usage Statement
Researchers may make free and open use of the UC Berkeley Library’s digitized public domain materials. However, some materials in our online collections may be protected by U.S. copyright law (Title 17, U.S.C.). Use or reproduction of materials protected by copyright beyond that allowed by fair use (Title 17, U.S.C. § 107) requires permission from the copyright owners. The use or reproduction of some materials may also be restricted by terms of University of California gift or purchase agreements, privacy and publicity rights, or trademark law. Responsibility for determining rights status and permissibility of any use or reproduction rests exclusively with the researcher. To learn more or make inquiries, please see our permissions policies (https://www.lib.berkeley.edu/about/permissions-policies).
