TL;DR
This paper proposes that eye fixations during simple choices are not random but follow an optimal strategy to gather information. By modeling this as an information sampling problem, they show that people tend to look at options that are both uncertain and have high estimated values, a pattern consistent with maximizing decision quality under cognitive constraints.
What Problem Does This Paper Address?
Making simple choices, like picking a snack from a shelf, involves more than just knowing what you like. We need to gather information about each option, and how we allocate our attention (where we look and for how long) plays a crucial role. Previous models explained how we integrate information (like the diffusion decision model) and how attention influences choices (e.g., fixating longer on an item makes it more likely to be chosen). However, a key question remained: what determines which item we fixate on and when? This paper tackles this by framing fixation as an optimal information sampling strategy, aiming to understand if our gaze patterns are driven by a rational process to make the best possible decision.
Methods
The authors developed a computational model based on optimal information sampling. They formalized the decision process as a metalevel Markov Decision Process (MDP), where "states" are beliefs about item values and "actions" are cognitive operations like sampling or stopping. Since finding the exact optimal policy is computationally intractable for more than two items, they used an approximation method based on features representing different types of information value (myopic, item-specific, and full information). This allowed them to derive a near-optimal policy that guides attention. They then fit this model to human choice and fixation data using an approximate maximum likelihood method.
Datasets
The study used two existing datasets:
- A binary food choice task [9].
- A trinary food choice task [10]. In both studies, participants rated snack items for liking (serving as a proxy for true value) and then made choices while their eye fixations were recorded.
Key Insights and Ideas
- Attention is not arbitrary: Fixations during simple choices are driven by an optimal strategy to gather information, not just random exploration or simple saliency.
- Value and Uncertainty are Key: The optimal policy dictates fixating on items that have both high estimated values and high uncertainty. This balances exploring potentially good options with confirming the value of already promising ones.
- Dynamic Allocation: Attention is not static; it's dynamically allocated based on evolving value estimates and their uncertainty throughout the decision process.
- Prior Beliefs Matter: Initial beliefs about item values (priors) play a significant role in shaping attention and choice, especially in explaining attention-choice correlations. A slightly biased prior can explain observed biases better than a neutral one.
- Binary vs. Trinary Differences: The model naturally captures differences in fixation patterns between binary and trinary choices, highlighting how the complexity of the choice set influences optimal sampling strategies.
Key Results
The model successfully accounted for many observed patterns in human choice and fixation data, including:
- Choice Curves: Accurately predicted how item ratings influence choice probabilities.
- Fixation Time Distributions: Provided a reasonable qualitative account of total fixation times.
- Fixation Duration: Captured key patterns like shorter final fixations and longer durations in binary choices compared to trinary.
- Uncertainty-Directed Attention: Showed that fixations are directed towards items with less cumulative fixation time (higher uncertainty), especially in trinary choices.
- Value-Directed Attention: Demonstrated a tendency to fixate on higher-rated items, reflecting the influence of estimated value.
- Attention-Choice Correlation: Explained the correlation between fixation time and choice, attributing it to both prior beliefs and value-directed attention.
Limitations and Future Work
- Approximation: The model uses an approximation for the optimal policy, which might not be perfectly accurate, especially for complex scenarios.
- Algorithmic Detail: The model focuses on the optimal policy rather than the precise algorithmic implementation in the brain. It doesn't specify how the brain actually computes these optimal strategies.
- Fixation Duration: The model underpredicts initial fixation durations in some cases.
- Single Fixation Trials: The model overpredicts single-fixation trials in binary choices.
- Future Directions: The authors suggest exploring more realistic constraints on fixation processes (e.g., inhibition of return), incorporating bottom-up influences (saliency), extending the model to multi-attribute choices, and investigating how the brain approximates optimal sampling policies.
Summary updated