Display Publications by [Year] [Type] [Topic]
This paper presents fitness critics as noise-reduced fitness evaluation functions for cooperative coevolution of multiagent teams. Noise in the evaluation function can have a negative impact on evolutionary learning, especially in multiagent learning. The evaluation function for an individual agent can be noisy as each agent in the team changes their policies independently. Existing noise-reduction methods are not suitable for efficient evolutionary algorithms. The noise- reduction provided by actor-critic methods cannot be directly applied to the evolutionary algorithms when learning policies in episodic problems as the critic provides stepwise evaluations but evolutionary algorithms require episodic fitness evaluations. Fitness critics can provide noise-reduced episodic fitness evaluations by aggregating the multiple stepwise evaluations of an intermediate critic. With demonstrations on the tightly coupled multi-rover domain, this paper shows that teams trained with fitness critics achieve comparable or increased team performance scores compared to teams trained with other noise-reduction methods.
(unavailable)
@InProceedings{tumer-rockefeller_mrs19,
author = {G. Rockefeller and P. Mannion and K. Tumer},
title = {Fitness Critics for Multiagent Learning (Poster session)},
booktitle = {IEEE International Symposium on Multi-Robot and Multi-Agent Systems (MRS)},
address = {Rutgers, NJ},
month = {August},
pages={},
abstract={This paper presents fitness critics as noise-reduced fitness evaluation functions for cooperative coevolution of multiagent teams. Noise in the evaluation function can have a negative impact on evolutionary learning, especially in multiagent learning. The evaluation function for an individual agent can be noisy as each agent in the team changes their policies independently. Existing noise-reduction methods are not suitable for efficient evolutionary algorithms. The noise- reduction provided by actor-critic methods cannot be directly applied to the evolutionary algorithms when learning policies in episodic problems as the critic provides stepwise evaluations but evolutionary algorithms require episodic fitness evaluations. Fitness critics can provide noise-reduced episodic fitness evaluations by aggregating the multiple stepwise evaluations of an intermediate critic. With demonstrations on the tightly coupled multi-rover domain, this paper shows that teams trained with fitness critics achieve comparable or increased team performance scores compared to teams trained with other noise-reduction methods.},
bib2html_pubtype = {Refereed Conference Papers},
bib2html_rescat = {Multiagent Systems, Evolutionary Algorithms},
year = {2019}
}
Generated by bib2html.pl (written by Patrick Riley ) on Wed Apr 01, 2020 17:39:43