Role of background knowledge in learning logical representations for embodied agents

Ascertain whether reliance on handcoded background knowledge is a limiting inductive bias for learning logical representations in embodied agents, or whether appropriate background knowledge should be identified and incorporated to support robust learning.

Background

Inductive Logic Programming and related symbolic methods often depend on background knowledge (predicates and initial theories) to learn rules. While effective in simplified domains, this reliance may hinder scalability and robustness in complex, noisy, physical environments.

The authors question whether background knowledge constitutes a limiting bias for embodied intelligence or if better identification of suitable background knowledge is necessary to enable learning compositional, lifted, and stochastic models for perception and action.