Language Representations for Generalization in Reinforcement Learning

Nikolaj S Goodger (Federation University Australia)*; Peter Vamplew (Federation University); Cameron Foale (Federation University); Richard Dazeley (Deakin University)

---

PMLR Page

Abstract

The choice of state and action representation in Reinforcement Learning (RL) can have a significant affect on the training task. But its impact on generalization is under-explored. One approach to improving generalization might be to use language as a representation. In this paper, we compare vector-states and discrete-actions to language representations and find that agents trained using language tend to generalize better than those using other representations. We also show how language models with extractive capabilities can allow generalization to new entities.