Abstract

Training reinforcement learning agents that continually learn across multipleenvironments is a challenging problem. This is made more difficult by a lack ofreproducible experiments and standard metrics for comparing different continuallearning approaches. To address this, we present TELLA, a tool for the Test andEvaluation of Lifelong Learning Agents. TELLA provides specified, reproduciblecurricula to lifelong learning agents while logging detailed data forevaluation and standardized analysis. Researchers can define and share theirown curricula over various learning environments or run against a curriculumcreated under the DARPA Lifelong Learning Machines (L2M) Program.

 

Quick Read (beta)