Systematic Generalisation through Task Temporal Logic and Deep Reinforcement Learning

This work introduces a neuro-symbolic agent that combines deep reinforcement\nlearning (DRL) with temporal logic (TL) to achieve systematic zero-shot, i.e.,\nnever-seen-before, generalisation of formally specified instructions. In\nparticular, we present a neuro-symbolic framework where a symbolic module\ntransforms TL specifications into a form that helps the training of a DRL agent\ntargeting generalisation, while a neural module learns systematically to solve\nthe given tasks. We study the emergence of systematic learning in different\nsettings and find that the architecture of the convolutional layers is key when\ngeneralising to new instructions. We also provide evidence that systematic\nlearning can emerge with abstract operators such as negation when learning from\na few training examples, which previous research have struggled with.\n