Periodic Reporting for period 1 - ACSAI (Automated Crowd Simulation with Artificial Intelligence)
Berichtszeitraum: 2023-09-01 bis 2024-08-31
The ACSAI project operates within the realm of digital transformation, specifically focusing on enhancing crowd simulation technologies used for planning major events and constructing large buildings. The core issue is that traditional planning methods are analog, slow, and prone to errors. Crowd simulation can be a key tool in addressing these problems, but existing simulation tools suffer from long computation times and limited flexibility, hindering their wide-scale adoption.
Our vision is to provide one-click simulation results that allow users to quickly assess and validate any design at a glance. Achieving this vision requires the use of advanced Artificial Intelligence (AI) techniques.
The main objective of ACSAI is to leverage advanced AI techniques,including Convolutional Neural Networks (CNN), to make crowd simulations faster, more realistic, and better integrated into early planning processes. During the course of the project, we realized that additional techniques, such as Reinforcement Learning (RL) and Dynamic Mode Decomposition (DMD), would also be necessary to reach our goal of seamless, real-time simulations.
By improving the efficiency of their crowd simulation software crowd:it, the project aims to enhance the safety, efficiency, and reliability of planning public spaces, events, and transportation hubs. This is especially relevant as urban areas expand, and public transport use is expected to double to meet European Green Deal climate targets.
Our vision is to provide one-click simulation results that allow users to quickly assess and validate any design at a glance. Achieving this vision requires the use of advanced Artificial Intelligence (AI) techniques.
The main objective of ACSAI is to leverage advanced AI techniques,including Convolutional Neural Networks (CNN), to make crowd simulations faster, more realistic, and better integrated into early planning processes. During the course of the project, we realized that additional techniques, such as Reinforcement Learning (RL) and Dynamic Mode Decomposition (DMD), would also be necessary to reach our goal of seamless, real-time simulations.
By improving the efficiency of their crowd simulation software crowd:it, the project aims to enhance the safety, efficiency, and reliability of planning public spaces, events, and transportation hubs. This is especially relevant as urban areas expand, and public transport use is expected to double to meet European Green Deal climate targets.
The ACSAI project made several advances in embedding AI technologies into their crowd simulation software. Besides the Convolutional Neural Networks (CNN) where we had first results from a previous research project, we have tested two more AI-based methods:
- Dynamic Mode Decomposition (DMD): This method was implemented to reduce the computational cost of dynamic flooding, the most resource-intensive part of the simulation. DMD was successfully integrated into the crowd:it software and applied to exemplary scenarios. Evaluations showed that it can provide a useful approximation of dynamic flooding, but there is a trade-off between accuracy and performance improvements. We have several ideas for improving this trade-off in order to leverage the full potential of this approach in enabling more scalable, high-speed simulations.
- Reinforcement Learning (RL): The RL approach aimed to improve the behavior of individual agents in the simulation, making them more adaptable and intelligent. Although the initial results showed promise, especially in terms of enabling agents to make real-time decisions based on their surroundings, further refinement is necessary to fully integrate this model into the product. Specifically, the team faced challenges in calibrating the RL model to handle the complexities of real-world scenarios with large numbers of agents, and this remains a focus for future research.
These technical advancements made possible through the ACSAI project ensure that the crowd:it software is evolving to meet the increasing demands for real-time crowd simulations in large-scale events and public infrastructure planning.
- Dynamic Mode Decomposition (DMD): This method was implemented to reduce the computational cost of dynamic flooding, the most resource-intensive part of the simulation. DMD was successfully integrated into the crowd:it software and applied to exemplary scenarios. Evaluations showed that it can provide a useful approximation of dynamic flooding, but there is a trade-off between accuracy and performance improvements. We have several ideas for improving this trade-off in order to leverage the full potential of this approach in enabling more scalable, high-speed simulations.
- Reinforcement Learning (RL): The RL approach aimed to improve the behavior of individual agents in the simulation, making them more adaptable and intelligent. Although the initial results showed promise, especially in terms of enabling agents to make real-time decisions based on their surroundings, further refinement is necessary to fully integrate this model into the product. Specifically, the team faced challenges in calibrating the RL model to handle the complexities of real-world scenarios with large numbers of agents, and this remains a focus for future research.
These technical advancements made possible through the ACSAI project ensure that the crowd:it software is evolving to meet the increasing demands for real-time crowd simulations in large-scale events and public infrastructure planning.
The ACSAI project delivered several innovations that position crowd simulation techiques, and in particular the software crowd:it, beyond the current state of the art.
The results with the CNN were too limited and the initial effort required to prepare a building model was too high for regular planning processes. This prompted us to test other AI techniques for more generalizable applications.
The ACSAI project has successfully contributed to advancing the state of the art in crowd simulations through the implementation of AI techniques.
DMD provides the most immediate perspective for improving our simulation software. If we manage to improve the communication between the model and our simulation kernel, it will significantly reduce the computational requirements of simulations by minimizing the need for repeated dynamic flooding recalculations. This will have a great impact, especially in scenarios requiring quick, reliable predictions, such as large events. The DMD method is integrated into the crowd:it software in a first version and has demonstrated its applicability to real-world scenarios, such as festival planning.
Reinforcement Learning (RL), while not yet fully market-ready, lays a strong foundation for future advancements. The RL model was designed to create more realistic agent behaviors, which is crucial for simulating crowd dynamics in complex and unpredictable environments. While further refinement is required, particularly in calibrating the agents' decision-making processes, the project has established a clear path for future development.
The results for both approaches—DMD and RL—are very promising, and each covers different application areas, giving us a broader use case beyond just station projects. These methods can also be applied to event venues and public spaces to ensure visitor safety.
Through ACSAI, we have enhanced the performance of crowd:it and laid the groundwork for future innovations. We are confident that this progress will lead to commercial expansion. For the next steps we will focus especially on refining the RL approaches and have integrated the AI tasks in our daily software development processes. Therefore, we have integrated the AI to-dos in our daily software engineering work. We plan to present a demonstrator to the beta testers of our software and gather feedback.
Additionally, we are well-connected within the scientific community of crowd simulation. The applicability of different AI methods in crowd simulation is still an emerging topic in the scientific field, and we actively contribute to this discussion. We present our ACSAI results at various conferences, from the PED (Pedestrian and Evacuation Dynamics) Conference to the Georg Nemetschek Institute Symposium on Artificial Intelligence for the Built World.
The results with the CNN were too limited and the initial effort required to prepare a building model was too high for regular planning processes. This prompted us to test other AI techniques for more generalizable applications.
The ACSAI project has successfully contributed to advancing the state of the art in crowd simulations through the implementation of AI techniques.
DMD provides the most immediate perspective for improving our simulation software. If we manage to improve the communication between the model and our simulation kernel, it will significantly reduce the computational requirements of simulations by minimizing the need for repeated dynamic flooding recalculations. This will have a great impact, especially in scenarios requiring quick, reliable predictions, such as large events. The DMD method is integrated into the crowd:it software in a first version and has demonstrated its applicability to real-world scenarios, such as festival planning.
Reinforcement Learning (RL), while not yet fully market-ready, lays a strong foundation for future advancements. The RL model was designed to create more realistic agent behaviors, which is crucial for simulating crowd dynamics in complex and unpredictable environments. While further refinement is required, particularly in calibrating the agents' decision-making processes, the project has established a clear path for future development.
The results for both approaches—DMD and RL—are very promising, and each covers different application areas, giving us a broader use case beyond just station projects. These methods can also be applied to event venues and public spaces to ensure visitor safety.
Through ACSAI, we have enhanced the performance of crowd:it and laid the groundwork for future innovations. We are confident that this progress will lead to commercial expansion. For the next steps we will focus especially on refining the RL approaches and have integrated the AI tasks in our daily software development processes. Therefore, we have integrated the AI to-dos in our daily software engineering work. We plan to present a demonstrator to the beta testers of our software and gather feedback.
Additionally, we are well-connected within the scientific community of crowd simulation. The applicability of different AI methods in crowd simulation is still an emerging topic in the scientific field, and we actively contribute to this discussion. We present our ACSAI results at various conferences, from the PED (Pedestrian and Evacuation Dynamics) Conference to the Georg Nemetschek Institute Symposium on Artificial Intelligence for the Built World.