Learning Parameterized Task Structures for Generalization to Unseen Entities

The biggest challenge AI faces when performing complex real-world tasks is understanding compositional task structure. Real-world tasks consist of multiple smaller subtasks combined in a specific order, and they also involve the sparse reward problem, where rewards are infrequent.

Existing research has not sufficiently considered the relationships between subtasks, resulting in a limitation where the task cannot be performed when a new entity not seen during training (e.g., a cabbage instead of an egg) appears.

PSGI (Parameterized Subtask Graph Inference), proposed by LG AI Research, uses a first-order logic approach to model the entire task through parameterized options and subtasks. This approach provides the following benefits:

• Increased efficiency: By using a parameterized format such as [Pickup x] to share common structures, redundancy is reduced and inference is possible with fewer samples.
• Zero-shot generalization: Even subtasks containing new entities not encountered during the training phase can be adapted to immediately if they share a similar structure.

Through experiments in the cooking, mining (Minecraft), and virtual AI2Thor domains, this study demonstrated that PSGI is more efficient and has stronger generalization performance than existing prior research.