Cluster errors

• General observations
  • NLL (negative log-likelihood) loss works better for convergence than MSE (mean-square error) loss
  • Scaling necessary (requires tuning of parameter -> currently only one parameter, but optimally take 2 separate ones)
  • Training now rather stable and testing different configurations
    • Also tested the idea sigma / sqrt(qTot). Worked decently well, reaches similar number of tracks, but efficiency is down by 10-15%
       
      • Feeding in both cluster and track position completely deteriorated the fit, no long tracks found
      • Getting rather good results now for long tracks even though total number of tracks is still not as high as with default method -> Need to adjust scaling parameter

• Next try: Retuning x and y scaling separately