Discovering Implied Serial Order Through Model-Free and Model-Based Learning.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Animal Behavior

Background:

• Humans and animals can learn arbitrary item orders without explicit cues.
• This learning appears to rely on transitive inference (TI), a fundamental property of ordered sets.

Purpose of the Study:

• To summarize research on the transitive inference paradigm.
• To compare the performance of six computational models in explaining TI behavior.
• To assess model-based versus model-free approaches in TI.

Main Methods:

• Review of existing research on transitive inference.
• Comparison of six computational models: three model-free (Q-learning, Value Transfer, REMERGE) and three model-based (RL-Elo, Sequential Monte Carlo, Betasort).
• Evaluation of models' ability to replicate empirically observed TI behaviors.

Main Results:

• Model-based computational approaches demonstrated better performance across a wider range of scenarios compared to model-free methods.
• No single computational model fully accounted for the entire spectrum of behaviors observed in the transitive inference literature.

Conclusions:

• Model-based approaches offer a more comprehensive framework for understanding transitive inference.
• Further development is needed to create a unified model that explains the full range of human and animal transitive inference behaviors.