Uncovering the computational mechanisms underlying many-alternative choice

Armin W. Thomas, Felix Molter, Ian Krajbich

TL;DR

When faced with many choices, we don't just pick the best option after looking at everything. Instead, our gaze actively influences how much we value items, and a model combining gaze-driven evidence accumulation with relative comparisons best explains how we make these complex decisions. This work bridges the gap between simple two-choice models and complex multi-choice scenarios, highlighting the active role of attention in decision-making.

What Problem Does This Paper Address?

We're constantly bombarded with choices, from what to buy at the supermarket to what to watch on streaming services. While research has explored how we make decisions between two options (often involving gaze-driven evidence accumulation) and how we might "satisfice" (settle for "good enough") when faced with many options, there's been a disconnect. This paper aims to bridge this gap by investigating the computational mechanisms behind many-alternative value-based choices, specifically exploring how gaze influences subjective value and decision strategies in scenarios with a large number of alternatives.

Methods

The researchers developed and compared several computational models of decision-making. These models varied in two key aspects:

  1. Decision Mechanism: They tested models based on:
  2. Optimal Choice: Assumes all items are examined, and the best is chosen.
  3. Satisficing: Assumes individuals stop searching once an item meets a threshold.
  4. Evidence Accumulation: Assumes evidence for each item is gathered and compared over time.
  5. Gaze Influence: They incorporated two accounts of how gaze affects decisions:
  6. Passive Gaze: Gaze only determines which items are considered.
  7. Active Gaze: Gaze actively increases the subjective value of attended items, either multiplicatively or additively.

They used eye-tracking data to record participants' gaze patterns while they made choices from sets of 9, 16, 25, or 36 snack food items. Model performance was evaluated using WAIC (Widely Applicable Information Criterion) for model fit and by simulating data to assess how well models predicted observed choices, response times, and the relationship between gaze and choice.

Datasets

The study used data collected from 49 healthy participants who made choices from sets of 9, 16, 25, or 36 snack food items. Participants also provided liking ratings for each item. No new benchmarks were introduced; the focus was on analyzing existing behavioral data through the lens of computational models.

Key Insights and Ideas

  • Gaze is Active, Not Passive: The most crucial finding is that models incorporating an active role for gaze (where looking at an item increases its perceived value) consistently outperformed models with a passive role. This suggests our eyes aren't just passive scouts but actively shape our preferences.
  • Satisficing Isn't Enough (Alone): While a probabilistic version of satisficing (PSM) with active gaze provided a good fit for choices and response times, it wasn't the best overall. This indicates that simply stopping at "good enough" doesn't fully capture the complexity.
  • Evidence Accumulation with Relative Comparison Wins: The gaze-driven evidence accumulation and comparison model (specifically, the GLAM model) provided the most comprehensive account, especially when considering the link between gaze allocation and the final choice. This highlights the importance of not just accumulating evidence but also comparing items relative to each other.
  • No Pure "Optimal Choice": Participants did not examine all items before choosing, ruling out a strict "optimal choice" model. They also didn't always choose the last item they looked at, disproving a simple hard satisficing strategy.

Key Results

  • Model Performance: Active gaze models significantly outperformed passive gaze models across all set sizes.
  • Best Model: The gaze-driven relative evidence accumulation model (GLAM+) provided the best overall account of the data, particularly in explaining the relationship between gaze and choice.
  • Probabilistic Satisficing: The probabilistic satisficing model (PSM+) offered a slightly better fit for choices and response times compared to other models.
  • Gaze Influence: Participants' choices were strongly influenced by the time they spent looking at each alternative, even after accounting for item ratings. This gaze influence increased with set size.
  • Response Times: Response times generally increased with set size, as expected.

Limitations and Future Work

  • Appetitive Choices: The study focused on appetitive snack foods. The authors suggest future work should investigate how these models apply to choices involving aversive options or reversed decision problems, where additive and multiplicative gaze biases might have opposite effects.
  • Familiarity and Learning: The experiment used familiar snack foods. The authors note that in novel environments, relative evidence accumulation might be more crucial than satisficing, and suggest future research explore model performance in novel vs. familiar choice environments.
  • Causality: While the study found strong correlations between gaze and choice, and between gaze and value, the authors acknowledge that the data support bidirectional causality (value influencing gaze, and gaze influencing value/choice), which is a complex area for future investigation.

Summary updated

Abstract

How do we choose when confronted with many alternatives? There is surprisingly little decision modelling work with large choice sets, despite their prevalence in everyday life. Even further, there is an apparent disconnect between research in small choice sets, supporting a process of gaze-driven evidence accumulation, and research in larger choice sets, arguing for models of optimal choice, satisficing, and hybrids of the two. Here, we bridge this divide by developing and comparing different versions of these models in a many-alternative value-based choice experiment with 9, 16, 25, or 36 alternatives. We find that human choices are best explained by models incorporating an active effect of gaze on subjective value. A gaze-driven, probabilistic version of satisficing generally provides slightly better fits to choices and response times, while the gaze-driven evidence accumulation and comparison model provides the best overall account of the data when also considering the empirical relation between gaze allocation and choice.
BibTeX
@inproceedings{thomas2021uncovering,
  author = {Armin W. Thomas and Felix Molter and Ian Krajbich},
  title = {{Uncovering the computational mechanisms underlying many-alternative choice}},
  booktitle = {eLife},
  year = {2021},
  doi = {10.7554/elife.57012},
}
Figure