Adaptive Mesh Refinement (AMR)

Our AMR software actively monitors the collection of high-dimensional data to ensure data is only collected where it increases the information about the space.  The algorithm uses the collected data to make predictions about the uncollected data.  If the difference between the predictions and actual collected data are smaller than a user-defined threshold, no more data is collected.  Additionally, we record all collected data so that future runs are even faster; there is no point in using time, money, or computational resources that you don't have to.

The animation below illustrates the naturally iterative process of AMR exploring a cognitive model parameter space.  The grey spheres indicate parameter combinations that are actually collected, while intersections on the grid indicate parameter combinations where a prediction was generated.  Ultimately, areas of the space determined to be linear are just predicted, while additional data is collected in areas where the data is non-linear.  The resulting surface has only twice the error that would be expected due to the stochastic nature of the model, despite being calculated with 1/100th of the computation!

We are pleased to announce that a limited-functionality distribution of our AMR software is nearly ready.  See below for some screen shots of the interface in action.