Record Details
Theoretical and implementation improvements for difference evaluation functions
ScholarsArchive at Oregon State University
| Field | Value |
|---|---|
| Title | Theoretical and implementation improvements for difference evaluation functions |
| Names |
Colby, Mitchell Kataichi
(creator) Tumer, Kagan (advisor) |
| Date Issued | 2014-04-02 (iso8601) |
| Note | Graduation date: 2014 |
| Abstract | Multiagent learning with cooperative coevolutionary algorithms is a critical area of research, and is relevant to many real-world applications including air traffic control, distributed sensor network control, and game-theoretic applications such as border patrol. A key difficulty in multiagent learning is the credit assignment problem, where the impact of each individual agent on the overall system performance must be ascertained. Difference evaluation functions aim to solve this credit assignment problem, by approximating the effect that each agent has on the system evaluation function. Difference evaluations have proven to produce superior learned policies in many multiagent settings. Although difference evaluations have produced excellent empirical results, there are still three key research questions that must be addressed regarding their usefulness in real-world systems. More specifically, the performance, theoretical advantages, and methodology for implementation must be addressed in order to demonstrate that difference evaluations are practical for use in real-world multiagent learning. These research questions are addressed in this dissertation. The first contribution of this dissertation is to demonstrate that difference evaluations may be extended and combined with other coordination mechanisms, resulting in superior learned performance. The second contribution of this dissertation is to derive conditions which guarantee that difference evaluations will outperform traditional coordination mechanisms. The third and final contribution of this dissertation is to demonstrate that difference evaluations may be approximated using only local knowledge, allowing for their implementation in any generic multiagent learning setting. By addressing the performance, theoretical foundation, and implementation concerns of difference evaluations, this dissertation provides a detailed analysis demonstrating the usefulness of difference evaluation functions in multiagent learning systems. |
| Genre | Thesis/Dissertation |
| Access Condition | http://creativecommons.org/licenses/by/3.0/us/ |
| Topic | Multiagent Learning |
| Identifier | http://hdl.handle.net/1957/47555 |
