Graduate Thesis Or Dissertation
 

Condensing observation of locale and agents : a state representation

Public Deposited

Downloadable Content

Download PDF
https://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/79408026s

Descriptions

Attribute NameValues
Creator
Abstract
  • Computing agents require state information to make coherent and useful decisions. A state representation is a numerical translation of the environment and conditions that are pertinent factors in an agent's decision making. Although many representations, when paired with clever learning algorithms, are able to coordinate and capture prey in specific domain types, an inherent problem is that they rarely scale well with domain changes. This is due to much of their information being of a form that is dependent upon factors such as the size of the world and number of agents. When the state information is instead a scaled and condensed view of surrounding agents, mapped to an action list from which the agent is able to choose, the state is then in a form that is independent of both world size and number of agents. Coupled with a simple, 1-step Q-Learning algorithm, this representation proves to be quite robust, outperforming a hand-coded policy and two other state-space representations. Using several 1v1 simulations, we will show that a predatory agent is capable of tracking and capturing a moving target in a simple gridworld domain. It is also able to transfer its experience to new domains that have larger gridworlds, more dynamic environments, and even sensor noise and failure.
License
Resource Type
Date Available
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Committee Member
Academic Affiliation
Non-Academic Affiliation
Subject
Rights Statement
Publisher
Peer Reviewed
Language
Replaces

Relationships

Parents:

This work has no parents.

In Collection:

Items

Thumbnail Title Date Uploaded Visibility Actions
file details ValentiBrettA2010.pdf 2017-08-01 Public Download