What Is the Model in Model-Based Planning?

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Human problem-solving exhibits remarkable flexibility, adapting to novel situations with minimal training.
• Existing research often focuses on cross-domain generalization, neglecting within-domain variation.
• Real-world tasks require generalizing across diverse within-domain changes.

Purpose of the Study:

• To investigate the role of representational abstraction in within-domain generalization.
• To explore how different task representations impact generalization in complex tasks.
• To compare agent and human generalization behaviors in novel game environments.

Main Methods:

• Developed a model-based planning framework to explore task representations.
• Tested how different classes of task representation influence generalization.
• Compared agent performance with distinct task representations to human performance in grid-based video games.

Main Results:

• Task representation significantly influences generalization behavior within a model-based planning framework.
• Agents using propositional rules (objects and relations) demonstrated effective within-domain generalization.
• Human performance in novel game tasks aligned with agents employing object- and relation-based representations.

Conclusions:

• Representational abstraction, particularly using propositional rules, is crucial for human within-domain generalization.
• This finding supports the hypothesis that human flexibility stems from abstract, rule-based task representations.
• The study provides a computational model for understanding human adaptability in complex problem-solving.