Minds for Mobile Agents

There is an increasing need to quantitatively describe and anticipate the way in which people walk around in public spaces (e.g., to maintain social distance, avoid collisions with automated vehicles etc.). The Predictive Pedestrian framework models the way in which people make step (direction and speed) decisions based on their spatial goals (e.g., to find items in supermarket) and priorities (e.g., to avoid other shoppers, stay near friends etc.). However, this fine-grained description of the cognitive states of simulated mobile agents is computationally demanding, making it difficult to apply in practice. This project efficiently implemented key calculations so that simulation of groups of pedestrians were sped up by an order of magnitude, achieving near real time for complex scenarios (e.g., shoppers in a crowded supermarket). A similar speed up was achieved in calculating the likelihood of model parameters based on observed movement trajectories. This will enable estimation of causes of individual differences in the way diverse people move around in different scenarios, underpinning more realistic simulations of pedestrian behavior that can inform and improve public policy and technological developments, such as the design of buildings and urban spaces and safer interactions with automated systems.

In January to March 2024, the project was extended through external funding. The extension focused on optimizing the data structures that are used in the pedestrian simulation and making the code more accessible and flexible. Moreover, changes in the parameter estimation algorithm led to further improvements reaching a speed-up between 10-20% compared to the previously optimized code.