Multi-Scale Reward Shaping via an Off-Policy Ensemble

TitleMulti-Scale Reward Shaping via an Off-Policy Ensemble
Publication TypeConference Paper
Year of Publication2015
AuthorsHarutyunyan, A, Brys, T, Vrancx, P, Nowe, A
Conference NameInternational Joint Conference on Autonomous Agents and Multi-Agent Systems (AAMAS-15)
PublisherTo appear
Abstract

We propose a potential-based reward shaping architecture that is able
to reduce learning speed, with no prior tuning and extra environment samples
required, via considering an off-policy ensemble of value functions
learning on a variety of heuristics with a variety of scales.