Honors College Thesis

 

Distortion of Agent States for Improved Coordination 公开 Deposited

可下载的内容

下载PDF文件
https://ir.library.oregonstate.edu/concern/honors_college_theses/pz50gx94d

Descriptions

Attribute NameValues
Creator
Abstract
  • Many real world problems have partial solutions or intermediary steps which can lead toward solving the problem. When assembling robotic teams to solve these problems, we have intuition about which intermediary steps are more useful than others. We examine methods to identify and apply our designer intuition onto tightly coupled multiagent problems. In a method analogous to potential based reward shaping, we shape the perceived value of points of interest (POI) in a ground-rover observation problem based on the potential for further team coordination if the observing agent goes to observe that POI. These state distortion methods utilize information from the current world, and as such are constructed independent of the learning method used. Methods for direct state shaping from Nasroullahi [1], which are based on future prediction about other agents’ actions, are extended into the future. From this initial work, we were inspired to create new methods that which readily scale to POI problems of arbitrary coupling dimension. These new methods show a no performance degradation in less coupled domains, and sustained operational capacity in more difficult, tightly coupled problems where traditional methods break down. This field of direct state distortion for increased cooperation and performance is relatively unexplored, and we finally lay out future directions of this area of work. Key Words: multiagent learning, state-shaping methods, tightly coupled problems
License
Resource Type
Date Available
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Non-Academic Affiliation
权利声明
Publisher
Peer Reviewed
Language
Replaces

关联

Parents:

This work has no parents.

属于 Collection:

单件