Thomas Schnürer, Malte Probst , Horst-Michael Groß, "Utilizing Emergent, Task-Independent Knowledge Representations for Accelerated Task-Learning in Reinforcement Learning", Fifth International Workshop on Intrinsically-Motivated Open-ended Learning, no. 5, 2022.

An intelligent agent in a complex environment will face a great number of diverse tasks. Rather than learning task-specific representations, we aim to reuse learned aspects to drive the acquisition of new tasks by leveraging previously learned abstract knowledge. Building on recent work that has introduced an inductive bias for explicit knowledge separation, we explore the benefits of such separation for learning new tasks. With an environment in which puzzle-like tasks can be created, we compare agents with different amounts of pre-trained knowledge and different ways of representing that knowledge.