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

Abstract

Users often struggle to locate an item within an information architecture, particularly when links are ambiguous or deeply nested in hierarchies. Information scent has been used to explain why users select incorrect links, but this concept assumes that users see all available links before deciding. In practice, users frequently select a link too quickly, overlook relevant cues, and then rely on backtracking when errors occur. We extend the concept of information scent by framing navigation as a sequential decision-making problem under memory constraints. Specifically, we assume that users do not scan entire pages but instead inspect strategically, looking "just enough" to find the target given their time budget. To choose which item to inspect next, they consider both local (this page) and global (site) scent; however, both are constrained by memory. Trying to avoid wasting time, they occasionally choose the wrong links without inspecting everything on a page. Comparisons with empirical data show that our model replicates key navigation behaviors: premature selections, wrong turns, and recovery from backtracking. We conclude that trial-and-error behavior is well explained by information scent when accounting for the sequential and bounded characteristics of the navigation problem.

Figures (12)

• Figure 1: Model architecture. The agent interacts with an external environment (a hierarchical information space, such as a menu) by choosing among three actions: visit, select, or return between pages to locate the target option. Each option has a latent information scent that becomes observable only upon inspection, and it is corrupted by perceptual noise. Observed cues are stored in working memory within the internal environment, where limited capacity and memory decay constrain retrieval. To handle uncertainty about where the target resides, the agent forms beliefs over states and selects actions according to its learned policy, receiving rewards (utility$-$cost) for reaching the goal and penalties for excessive steps. This resource-rational formulation frames human trial-and-error navigation as a sequential decision-making problem solvable with reinforcement learning.
• Figure 2: Illustration of the model’s decision process and state of working memory in a two-level menu task. The agent navigates by choosing among visit, select, and return actions, guided by perceived information scent (darker green = stronger cue). Each option has a latent scent; visits reveal noisy observations that update the agent’s belief. Scent is stored in working memory at two scopes: the top row tracks the global top-$k$ memory ($k=3$ here) aggregated across visited pages; the bottom row encodes the currently observed page. (a–b) On the top-level menu, visits expose option scents; the bottom row darkens for inspected items while the top row retains the strongest $k$ cues across panels. (c) A strong local cue prompts select (e.g., “Invertebrate Animals”), entering level two. (d) Further visits sample suboptions, sharpening local beliefs. (e) If local scents are weak, the agent may return; and with memory decay, entries dropping below the threshold are forgotten ($\times$), so the agent forgets having visited the two squares. (f–h) The agent continues visiting and selecting the final target.
• Figure 3: Overall parameter sensitivity across three perturbation levels (5%, 10 25%). Aggregated sensitivity values reflect the combined deviation in the three target effects when perturbing each parameter. The overall trend demonstrates that the model exhibits stable behavior under small perturbations and predictable, increases in sensitivity as perturbation levels grow.
• Figure 4: Navigation performance across task difficulties. Human data (top) show that both solution time and click count increase systematically from no-problem to competing to low-scent conditions. Our model (bottom) reproduces this gradient, with more steps and clicks required under greater uncertainty. Beyond matching overall performance, the model also captures human-like behavioral dynamics: direct paths in no-problem tasks, trial-and-error loops in competing tasks, and revisits under low-scent conditions where weak cues and memory decay hinder efficient navigation.
• Figure 5: Effect of hierarchy depth on navigation performance. Human data (top) shows that three-level hierarchies (8×8×8) significantly increased solution time and lostness compared to two-level hierarchies (16×32). Using our layouts with the same number of targets (bottom), the model reproduces this depth effect: three-level structures (4×4×4) required more steps and yielded higher lostness scores than two-level structures (8×8). This alignment indicates that deeper hierarchies amplify memory demands and trial-and-error exploration, making navigation less efficient for both humans and the model.