McGill University develops a three-agent workflow to collect data and forecast mobility choices under weather fluctuations

A study by McGill University researchers Narges Ahmadi, Youbo Jiao, Jonatas Augusto Manzolli, Jiangbo Yu, and Luis Miranda-Moreno presents a three-agent framework that directly connects data collection, mobility behavior, and weather-sensitive travel demand forecasting. The architecture aims to bridge a longstanding divide in travel behavior research, where digital data collection, predictive modeling, and evaluation have traditionally been developed and assessed in isolation.

The proposed workflow coordinates conversational data collection, structured data processing, and behavioral forecasting. For data collection, the team deployed an image-supported interactive chatbot to run a stated-preference survey on commuting mode choices among university students across five preset weather scenarios, yielding 454 respondent-scenario observations.

To analyze weather-related patterns and establish solid baselines, the researchers used a multinomial logit model representing conventional discrete choice analysis alongside standard machine learning algorithms, including logistic regression and random forest. The random forest model reached 69.6 percent accuracy across the five travel choices, serving as the primary performance baseline.

The empirical evaluation examined nine locally deployed large language models ranging from 2 billion to 35 billion parameters, comparing them across four zero-shot prompt and context configurations. Subsequent testing expanded to persona-based setups, few-shot prompting, and vision-based configurations. The top text-only language model achieved 69.9 percent accuracy in a zero-shot setting without any task-specific fine-tuning.

Analysis of prompting strategies showed that providing models with habitual commute information yielded the most consistent gains in prediction accuracy, while expert prompting generally outperformed persona role-playing. Personality trait data proved most useful when habitual commute details were unavailable. Few-shot prompting also improved predictions across several models, though performance gains plateaued after only a few examples.

Under vision-based configurations using the same weather images presented to survey respondents, the top multimodal model reached 71.5 percent accuracy in the five-way choice prediction, confirming that visual context provides certain models with useful predictive signal. The study concludes that conversational surveys, structured data processing, discrete choice modeling, machine learning, and multimodal language model forecasting can be successfully orchestrated within a fully auditable multi-agent system.