Nitschke, Geoff and Didi, Sabre (2018) Policy Transfer Methods in RoboCup Keep-Away, Proceedings of Genetic and Evolutionary Computation Conference (GECCO 2018), Kyoto, Japan, 117-118.
![[img]](http://pubs.cs.uct.ac.za/style/images/fileicons/application_pdf.png) |
PDF
2018-Policy_Transfer_Methods_in_RoboCup_Keep-Away.pdf Download (374kB) |
Abstract
This study investigates multi-agent policy transfer coupled with behavior adaptation by objective and non-objective search variants of HyperNEAT in RoboCup keep-away. For comparison, evolved behaviors were compared to those adapted by RL methods: SARSA and Q-Learning, coupled with policy transfer. Keepaway was selected as it is an established multi-agent experimental platform. Similarly, the SARSA and Q-Learning methods were selected as both have been demonstrated for boosting behavior quality with policy transfer. Keep-away behaviors were gauged in terms of effectiveness and efficiency. Effectiveness was average task performance given policy transfer, where task performance was average ball control time by the keeper team. Efficiency was average number of evaluations taken to reach a minimum task performance threshold given policy transfer.
| Item Type: | Conference poster |
|---|---|
| Subjects: | Computing methodologies > Artificial intelligence |
| Date Deposited: | 09 Nov 2018 |
| Last Modified: | 10 Oct 2019 15:31 |
| URI: | http://pubs.cs.uct.ac.za/id/eprint/1283 |
Actions (login required)
 |
View Item |