Modeling Trial-and-Error Navigation With a Sequential Decision Model of Information Scent

Information scent is a central construct in Information Foraging Theory, describing how users perceive the value, cost, or access path of information sources from proximal cues such as link descriptors, images, or page arrangement. Early studies, like those by Pirolli and Card, suggested that users should leave a page once its perceived yield drops below that of alternatives. However, this idealized navigation behavior rarely occurs in practice. Users often do not evaluate all options, sometimes commit prematurely, and revisit earlier pages. These behaviors suggest that information scent guides decisions under bounded rationality, operating through partial and uncertain evidence rather than fully rational strategies. Existing computational models, such as CoLiDeS and SNIF-ACT, provide some solutions but typically assume users can fully observe all options, overlooking the impact of memory constraints and uncertainty on decision-making.

Navigating complex information architectures is a challenging problem, particularly when links are ambiguous, overlapping, or deeply nested in hierarchies. Existing models often assume users can fully observe all options and make myopic decisions at the current page. However, users frequently make premature selections, overlook relevant cues, and rely on backtracking when errors occur. How to simulate user navigation behavior under memory constraints and uncertainty remains an unsolved problem.

The core innovations of this paper include extending information scent theory into a sequential decision framework, proposing a navigation model based on partially observable Markov decision processes (POMDP). 1) The model frames navigation as a sequential decision problem under memory constraints, where users make decisions based on local and global information scent within their time budget. 2) It introduces memory decay mechanisms to simulate the limitations and uncertainties of human memory. 3) The model employs reinforcement learning strategies to optimize navigation paths under limited cognitive resources. These innovations not only fill theoretical gaps in existing models but also offer more accurate predictions of user behavior in practice.

• �� Information Scent Calculation: Use a pre-trained sentence transformer to compute the semantic similarity between option labels and the goal. • Memory Decay Mechanism: Simulate the limitations and uncertainties of human memory, restricting the number of cues that can be retained and applying decay so that unrehearsed traces fade over time. • Reinforcement Learning Strategy: Employ reinforcement learning strategies to optimize navigation paths under limited cognitive resources. • Partially Observable Markov Decision Process (POMDP): Frame the navigation problem as a POMDP under memory constraints, where users make decisions based on local and global information scent within their time budget.

The experimental design includes using reconstructed HTML menu materials to simulate navigation tasks with varying task difficulties and hierarchy depths. The benchmark dataset used includes the Blackmon dataset, testing 9 headings and 93 links. Experimental metrics include solution time, click count, success rate, first-click accuracy, and lostness. Ablation studies confirm the contributions of memory decay and noise components to navigation efficiency and human-model alignment.

Experimental results show that the model effectively simulates user trial-and-error behaviors, replicating premature selections, wrong turns, and backtracking. These behaviors align closely with empirical user data, demonstrating the model's validity. Additionally, the model accurately captures user navigation patterns in experiments with varying task difficulties and hierarchy depths, such as increased error rates and solution times with higher task difficulty. Ablation studies confirm the contributions of memory decay and noise components to navigation efficiency and human-model alignment.

The model can be directly applied to user interface design, helping designers optimize information architectures to reduce user navigation errors and improve user experience. In an era of information overload, this model helps improve user experience by reducing navigation errors. Additionally, the model can be used in educational settings to help students navigate complex information architectures more effectively.

The model may underperform in handling extremely complex navigation structures due to memory constraints and noise leading to decision errors. Additionally, the model's parameters require tuning for different application scenarios, potentially limiting its generalizability. Future research could explore applying this model in more complex navigation environments or integrating it with other cognitive models to improve predictive accuracy.