Tasks

• Improve regression NN, check performance on simulation (peak width of network to ideal assignment and native to ideal assignment)
• Simulate clusters and check network CoG individually

Improve regression NN and peak width

• Removing the output qIdeal for the prediction strongly improved regression performance - Network only learns position of CoG for now

• Bug-fix which assigns the native clusters and network clusters correctly now (previous result was not accurate - wrong index assignment)

Simulated clusters

• Simple gaussian clusters are generated
• For purpose of simplicity only 2 within the 11x11 grid that the network gets
• First cluster mean between [-0.5, 0.5] for pad and time, second cluster, random within a radius of [-3,3] in the grid
• Checking position of the mean from which the gaussian was generated to the prediction of the NN and the CoG predicted by a dummy clusterizer (takes only the innermost 9 charge values and performs a CoG calculation)

• The following diagrams show the deviation of the CoG - mu of the central gaussian (left: network, right: native dummy CoG, x: time, y: pad)

• Network shows a clearly sharper peak in the middle, less spread and no outliers at high pad and time

To-Do / Next steps

• Optimize classification network -> This will be a more crucial part as the native clusterizer is already performing very well for regression (2D)
• Switch to 3D input and check whether this improves performance