Retrospective Inference as a Form of Bounded Rationality, and Its Beneficial Influence on Learning.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Cognitive models often use Bayesian filtering, combining new data with fixed past beliefs.
• This approach is computationally efficient but leads to suboptimal beliefs about past states.
• Knowledge of past states is valuable for intrinsic reasons and for learning environmental parameters.

Purpose of the Study:

• To introduce a novel computational approach for cognitive agents that balances accuracy and resource costs.
• To address the limitations of Bayesian filtering and Bayesian smoothing in real-time inference.
• To propose a method that improves both inference and learning in probabilistic cognitive models.

Main Methods:

• Proposed finite retrospective inference (FRI), a form of online fixed-lag smoothing with a sliding window.
• Developed a simple variational scheme for updating beliefs about a limited number of past states.
• Simulated the approach using randomly generated Hidden Markov Models (HMMs) and the probabilistic reversal task.

Main Results:

• FRI significantly increases the accuracy of inference about past states.
• The method enhances learning of fixed or slowly changing environmental parameters.
• Simulations demonstrated the effectiveness of FRI in HMMs and a specific probabilistic task.

Conclusions:

• Finite retrospective inference provides a computationally tractable solution for boundedly rational agents.
• FRI offers a theoretical contribution to normative accounts of bounded rationality.
• The approach yields testable empirical predictions for future research in cognitive modeling.