Different types of sampling: less&long vs. many&short simulations
There can be two types of sampling distinguished. Assuming the general case of using the data for a data-driven method, such as the surrogate model or UQ + time-dependent QoI:
- sample a (smaller) set of states (think: manifold) and run the simulation long. This gives less but long trajectories over time.
This case is simpler to do in VADERE, because the trajectories usually go from the "intuitive" start to end. Usually, the trajectory terminates when the scenario "is done" (e.g. all peds. evacuated). In the operator view, this is more suited to push forward functions. This is basically already supported.
- sample a (larger) set of states (think: manifold) and run short simulations. This gives many trajectories but short over time.
This is harder to carry out in VADERE, because trajectories need to start not only at the "intuitive start" (i.e. when the people get spawned) but start simulations during all times of scenario. This requires to set up a state in VADERE which, for example, requires to set the pedestrians in a meaningful way... This way of sampling is usually more suitable to push densities over time, and therefore, UQ. Intuitively, the short trajectories give a vector field on the state space and indicate in which direction the system would go (if it is in that state).
- Evaluate whether point 2 is required (start with small examples where point 2 is not too hard to accomplish, then see how much impact the different sampling method have).
- Think about: Can this sampling be replaced by a surrogate model? Is this only required on a macroscopic level?
- What are ways to integrate this in the suq-controller? There is much more controllability required for the initial state of a simulation.