Investigating Transfer Learning in the Urban Combat Testbed.

摘要

We have developed a complex agent capable of transferring learned declarative knowledge between tasks in order to improve performance in a rich domain developed to test transfer performance 'the Urban Combat Testbed'. We present and discuss our experiments and results which demonstrate successful transfer learning. To better understand the underlying sources of transfer in this domain, we empirically evaluate the types of knowledge that contribute to the observed improvement in performance due to transfer, and discuss the implications of our analysis for general transfer learning research.