Recent advances in decision neuroscience have increasingly focused on foraging-based studies to explore the underlying mechanisms of decision-making. Building on this research, we applied a well-established patch-foraging framework to investigate human decision-making in foraging contexts. Our findings replicated earlier results, demonstrating that human foragers deviate from optimal models, such as the Marginal Value Theorem (MVT).
To address the deterministic limitations of MVT, we explored stochastic action-selection algorithms, such as epsilon-greedy, softmax, and mellowmax. These models were assessed based on their ability to simulate foraging dynamics, incorporating key factors like the patch effect and environmental influences.
Our analysis revealed the following:
- The softmax model required an additional bias term to capture individual differences.
- The mellowmax model, with adaptive parameters, successfully modeled the diverse foraging behaviors.
Moreover, we incorporated methods to account for uncertainty in reward decay rates, providing insights into how uncertainty drives stochasticity in patch-leaving decisions.
We implemented a patch-foraging framework to study human decision-making processes in various foraging tasks. The experimental design built upon Le Heron’s study, which provided key insights into patch-leaving behavior across different environments.
Figure 1: Le Heron Patch Foraging Experimental Setup
Our analysis explored patch-leaving behavior under different environmental conditions. The variation in behavior was examined across different patch types, reflecting how individuals deviate from the Marginal Value Theorem.
Figure 2: Patch-Leaving Behavior Across Patch Types and Environmental Conditions
To model the observed deviations from optimal decision-making, we applied stochastic algorithms such as:
- Epsilon-Greedy Algorithm: Exploring the effect of varying epsilon values on leave times.
- Softmax Algorithm: Evaluating decision-making with probabilistic action selection.
- Mellowmax Algorithm: A smoother alternative to softmax, capturing a broader range of foraging behaviors.
Figure 3: Mean Leave Time Across Patch Types for Different Epsilon Values
Figure 4: Evaluation of Softmax-Based Models
Figure 5: Evaluation of Mellowmax-Based Models
To account for individual differences in foraging behavior, we fitted both softmax and mellowmax models to the subject-specific data. The mellowmax model, with its adaptive parameters, proved to better capture variability across different environmental conditions.
Figure 6: Comparison of Softmax and Mellowmax Parameter Fits to the Empirical Data
Human decision-making in foraging contexts exhibited significant variability in leave times. By analyzing subject-specific data, we observed how both environmental conditions and individual preferences shaped decision strategies.
Lastly, we introduced methods that integrate uncertainty into reward decay rates, providing a fresh perspective on how uncertainty contributes to stochastic patch-leaving decisions. This approach offers valuable insights into how humans balance exploration and exploitation in uncertain environments.
Figure 10: Exploration and Uncertainty-Driven Stochasticity in Patch-Leaving Decisions
Our study provides compelling evidence that stochastic models, particularly the mellowmax approach, offer valuable improvements over deterministic models like the MVT when analyzing human foraging behavior. By incorporating both individual differences and environmental uncertainty, these models give a more nuanced view of decision-making strategies in complex, dynamic environments.
For further information, questions, or collaboration inquiries, feel free to reach out over email or Linkedin.