Using deep neural networks as a guide for modeling human planning.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Developing cognitive models often involves refining initial intuitions against behavioral data.
• Distinguishing random variation from meaningful model deviations is a significant challenge in cognitive modeling.
• Deep neural networks (DNNs) offer a powerful tool for pattern detection in large datasets.

Purpose of the Study:

• To apply model comparison with DNNs to enhance a heuristic search model of human planning.
• To systematically identify opportunities for cognitive model improvement using large-scale behavioral data.
• To investigate the feasibility of scaling cognitive models to massive datasets for sequential decision-making research.

Main Methods:

• Developed a heuristic search model for human play in the combinatorial game 4-in-a-row.
• Trained deep neural networks on a dataset of 10,874,547 human games to predict moves.
• Compared the heuristic model against the performance of the trained DNNs to identify discrepancies.

Main Results:

• Deep neural networks accurately predicted human moves in 4-in-a-row, capturing meaningful behavioral patterns.
• Deviations between the heuristic model and DNNs highlighted specific areas for model refinement.
• Three extensions were added to the heuristic model, including opening biases and endgame planning adjustments.

Conclusions:

• Model comparison with high-performance DNNs is advantageous for cognitive model development.
• This approach enables systematic refinement of cognitive models even after initial testing.
• Scaling cognitive models to massive datasets is feasible and effective for understanding human sequential decision-making.