Calibrating intelligent, diverse agents to help predict real-world pedestrian behavior.
Together with the Amsterdam Mathematical Psychology Laboratory, HUBS initiated the Minds For Mobile Agents (M4MA) project, centered around pedestrian models. Using tools from cognitive decision modeling and behavioral data science, M4MA aims to understand the movement patterns of diverse groups of pedestrians as they would occur in real life.
Calibration of diverse agents included variables such as preferred speed and interpersonal distance. Image credits: HUBS
In the “predictive pedestrian” model, simulated agents are equipped with a basic Theory of Mind. This enables them to predict the behavior of other agents and base their own movement direction on those predictions. The mind of each agent is specified by parameters that have clear psychological interpretations, such as their preferred speed and interpersonal distance, and an agent’s decisions are governed by a utility-maximizing framework that is easily expandable. In addition, each agent can have their own set of goals that they need to fulfill within the space, and the ability to strategically plan their routes. In this manner, simulated pedestrians are able to navigate complex spaces and obtain desired targets autonomously, as for example in a supermarket scenario.
Diverse agents moving through a space. Image credits: M4MA
Movement data from real-life experiments enable us to calibrate the model to improve our simulations in terms of accuracy and predictive power. We are currently working on methods to estimate the model’s parameters from data. Additionally, other future developments of M4MA will enable the model to account for individual differences through Bayesian hierarchical modeling. While most pedestrian models focus on highly crowded scenarios and agents with simple goals, M4MA’s strength lies inmodeling lower-density pedestrian flows involving complex goals and individual differences.
Read more about M4MA here or check out the student projects of Gigi Vissers and Jonne Zomerdijk (winner UvA Thesis Award 2021 ).