Learning to Decompose: Human-Like Subgoal Preferences Emerge in Neural Networks Learning Graph Traversal

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Machine Learning

Background:

• Human problem-solving involves partitioning complex tasks into smaller subgoals.
• Existing research identifies principles governing human subgoal preferences but not their origin or efficiency.
• The effectiveness and efficiency of human subgoal preferences remain incompletely understood.

Purpose of the Study:

• To investigate the implicit learning processes underlying human subgoal preferences.
• To determine if learning optimal traversals can lead to human-like path decomposition in artificial agents.
• To explore the origins of effective and efficient problem-solving strategies.

Main Methods:

• Utilized graph-based environments and neural networks as learning models.
• Trained transformer models to learn shortest-path traversals.
• Evaluated models on known and novel graphs for unseen problems.

Main Results:

• Transformer models developed a preference for frequently occurring nodes on shortest paths, mirroring human subgoal preferences.
• This preference emerged even for unseen problems in new graphs.
• The same preference did not arise from learning random or Hamiltonian traversals.

Conclusions:

• Human-like subgoal preferences can implicitly emerge through learning shortest-path traversals.
• Explicit preference computation or exhaustive search is not necessary for developing these strategies.
• The findings highlight the role of data distribution in shaping learned problem-solving behaviors across different neural network architectures.