K&C seeks to model and understand extreme environments. Many of these environments present challenging problems which are not always solvable by direct methods. One such area is in the field of human crowd modeling. Crowds can exhibit highly non-linear behavior, which only is exasperated during times of panic. While the underlying physic-based equations might not be known, data exists that capture a part of the full picture.
K&C is exploring how this data, along with reinforcement learning might be better able to predict and model these complex phenomena. Through developing a novel pipeline, positional data can be digested and trained upon to create agents whose behavior is a generalization of the data. Built on top of a state-of-the-art gaming engine, these trained agents can then be deployed in new environments.
While the direct data only provides position and time, agents after a full training, exhibit generalized behavior such as obstacle avoidance.
The video demonstrates the process of training and illustrates the agent actively learning.
Next Steps
K&C continues to support this effort through the following:
- Continue sourcing relevant datasets.
- Identify possible collaborators.
- Aim for use in production software in 2022.
For More Information