First iteration
Problem: We first tried deep reinforcement learning to find potent paths. This takes each cell as an agent and let it learn paths toward EMT autonomously. However, we later gave up this method as reinforcement learning methods are quite hard to converge. Meanwhile, the results from deep learning networks are hard to explain, which also motivated us to look for another method.
Solution: We finally chose traditional graph-based path finding method, which is easier to perform and explain.
Second iteration
Problem: We originally performed the algorithm in PCA space, trying to find a path based on Euclidean distance. This method eventually failed we could not find a steady path, which is shown in the figure below.
Solution: We chose to made use of a Gaussian kernel to measure the distance between sample points in the dataset.
summary
Each engineering algorithm needs several iterations to reach best performance. We select two of them for presentation here, and we will continue to refine our metric in the future.