Determining social mechanisms for sequential decision-making in a virtual pedestrian route choice experiment

Moving groups are routinely faced with a choice of different routes as part of their daily lives, such as choosing between exits from a building. Differences in moving speeds and environmental constraints often lead to individuals being able to observe the choices of others ahead. This social information can inform their decision-making, but exactly how this is being used remains an open question. Previous theoretical studies on animal groups have demonstrated that simple heuristics are plausible and accurate mechanisms, with some further predicting that more recent decisions are more heavily weighted. Experiments with fish corroborate the importance of more recent decisions; however, experimental work is limited. Here, we conduct an online survey with human participants to identify which social decision-making mechanism individuals follow. Contrary to previous experimental work, we find little indication that recent decisions are weighted more heavily; instead, our results suggest that following the majority of previous decisions is the dominant behaviour. Furthermore, self-reported decision-making mechanisms correlate with our experimental findings despite their variability, suggesting that on average individuals can recognize their behavioural tendencies. Our findings give insight into social sequential decision-making, and provide an empirical foundation for integrating realistic social decision mechanisms into pedestrian movement models.

💡 Research Summary

**
This paper investigates how humans use socially transmitted information when making sequential binary route‑choice decisions in an evacuation‑type scenario. Building on a theoretical framework that models the probability of choosing option A as

(P_A = \frac{1}{1 + a,s^{-I}})

where a represents non‑social (personal) information, s captures the strength of social influence, and I encodes the social cue, the authors examine three concrete implementations of I: (1) an aggregate strategy that uses the raw difference in the number of prior choosers (Δn), (2) a majority strategy that reduces Δn to its sign (i.e., which option has the majority), and (3) a dynamic strategy that considers only the most recent decision (d). A fourth “dependencies” strategy that would require remembering the entire ordered sequence is omitted as implausible for human cognition.

To test which of these heuristics best describes actual human behaviour, the authors conducted an online experiment using a desktop‑based virtual environment built in Unity 3D. Participants were told they were evacuating a building and were shown a short video of eleven computer‑generated pedestrians moving down a corridor and choosing between two unmarked exits (A = left, B = right). The first nine pedestrians followed a fixed pattern (ABBABAABA) that slightly favoured exit A. The last two pedestrians were varied across four conditions:

1. AAA – both the majority and the most recent decision favour A.
2. ABA – majority favours A, most recent decision also A, but a B precedes it.
3. AAB – majority favours A, but the most recent decision is B.
4. ABB – both majority and most recent decision favour B.

After viewing the video, participants selected their own exit and then answered two open‑ended questions: why they chose that exit and whether they noticed any difference between the two videos (a second video was later shown to probe learning effects). Demographic data (age, sex) and response times were also recorded.

The behavioural data were fitted to the three candidate models using maximum‑likelihood estimation, and model comparison was performed with log‑likelihood values and the Bayesian Information Criterion (BIC). The majority strategy consistently achieved the lowest BIC, indicating the best trade‑off between fit and complexity. The aggregate strategy performed similarly but did not improve the fit enough to justify its extra parameterisation. In contrast, the dynamic strategy—where only the most recent decision influences the participant—failed to explain the data; participants frequently chose the option with the overall majority even when the latest observed decision pointed to the opposite side (conditions AAB and ABB).

Qualitative analysis of the open‑ended responses revealed that participants often explicitly referenced the number of people choosing each exit (“most people went left”) or the perception that “the crowd was heading that way,” confirming that self‑reported decision heuristics aligned with the majority‑following behaviour identified statistically. Moreover, the second video, which altered the final three decisions, did not produce systematic changes in participants’ choices, suggesting that the observed pattern reflects a stable heuristic rather than a short‑term learning or habituation effect.

The authors discuss these findings in relation to prior work on animal groups, notably studies on zebrafish that reported a strong weighting of the most recent decision. They argue that humans, perhaps due to higher cognitive load or a preference for robust cues in high‑stakes contexts, rely on a simpler “follow the majority” rule. This rule reduces the need to track the exact order of past decisions and provides a quick, low‑effort heuristic that still yields socially coordinated outcomes.

From an applied perspective, the results have direct implications for pedestrian simulation models, especially those used for emergency evacuation planning. Many existing crowd‑simulation frameworks incorporate social influence as a static parameter (e.g., proportion of agents at each exit) but often ignore the temporal ordering of observed choices. The empirical evidence presented here supports the inclusion of a majority‑following rule as a realistic behavioural component, which could improve the predictive accuracy of evacuation models and inform the design of signage or communication strategies that deliberately shape perceived majority behaviour.

The paper also acknowledges limitations of the virtual‑experiment approach: while desktop simulations capture basic route‑choice dynamics, they lack the full sensory and emotional richness of real emergencies. Consequently, the authors recommend future work that validates the majority‑following heuristic with field data or immersive virtual‑reality setups, and that explores how additional cues (e.g., signage, perceived safety, crowd density) interact with the majority effect.

In summary, the study provides robust experimental evidence that, in sequential decision‑making contexts, humans preferentially weight the overall majority of prior choices rather than the most recent observation. This insight bridges theoretical models from collective animal behaviour with practical crowd‑dynamics modelling, offering a concrete, empirically grounded mechanism for enhancing the realism of pedestrian evacuation simulations.