Meta-Referential Games to Learn Compositional Learning Behaviours

Kevin Yandoka Denamganai, Sondess Missaoui, James Alfred Walker

Keywords: language emergence, language grounding, compositionality, systematicity, few-shot learning

Abstract: Referring to compositional learning behaviours as the ability to learn to generalise compositionally from a limited set of stimuli, that are combinations of supportive stimulus components, to a larger set of novel stimuli, i.e. novel combinations of those same stimulus components, we acknowledge compositional learning behaviours as a valuable feat of intelligence that human beings often rely on, and assume their collaborative partners to use similarly. In order to build artificial agents able to collaborate with human beings, we propose a novel benchmark to investigate state-of-the-art artificial agents abilities to exhibit compositional learning behaviours. We provide baseline results on the single-agent tasks of learning compositional learning behaviours, using state-of-the-art RL agents, and show that our proposed benchmark is a compelling challenge that we hope will spur the research community towards developing more capable artificial agents.

One-sentence Summary: Presentation of a novel benchmark to study artificial agents' abilities to learn compositional learning behaviours, along with results of state-of-the-art RL algorithms against it.

Community Implementations: [ 2 code implementations](https://www.catalyzex.com/paper/arxiv:2207.08012/code)