Implicit Value Updating Explains Transitive Inference Performance: The Betasort Model.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Animal Behavior

Background:

• Transitive inference is a key aspect of serial learning across many species.
• Standard reinforcement learning models struggle to replicate transitive inference abilities observed in behavior.

Purpose of the Study:

• To introduce betasort, a novel algorithm for performing transitive inference.
• To compare betasort's performance against established models and biological subjects.

Main Methods:

• Betasort represents stimuli using beta distributions and updates all stimuli per trial.
• Asymmetric feedback processing and cognitive inspiration are key algorithmic features.
• Performance compared across rhesus macaques, humans, and Q-learning (a reward-prediction error model).

Main Results:

• Betasort successfully performed transitive inference.
• Q-learning models failed to perform above chance on critical test trials.
• Betasort demonstrated computational efficiency compared to full Markov decision process models.

Conclusions:

• Betasort offers a computationally efficient, cognitively inspired model for transitive inference.
• Reinforcement learning research benefits from feature-driven comparisons of formal models.
• This work highlights limitations of current reward-prediction error models in explaining complex inference.