Requirements for the Loss and Training Visualizer

Overall Requirements Specification

Trainer Module

• Implement QAOA for a bunch of problems
  • Max Cut
  • Weighted Max Cut
  • Travelling Salesman Problem
  • Multi-source Travelling Salesman
• Implement Methods of Parameter Searches
  • Gradient Descent
  • Full Grid Scan (for small p)
• Write a convenient trainer and evaluator
  • To find the best parameters
  • To get the loss history over the training
  • To test against new data and contrast against classical results (for small problems)

Loss Visualization Module

• Compute the Minimizers
  • Get multiple different minima (Random shot, Start from one and search other)
  • Find Maximizers of the same surface
  • (?) Try to compute Saddle Points
• Select Subspace to plot
  • Take points of interest (2-3 minimizers) and compute the subspace they lie in
  • Take random vectors to sample a random subspace
  • (?) Take random vectors orthogonal to already sampled subspace
• Make Different Plots
  • 1D plot of full scan from one point to another
  • 2D contour plot of full scan
  • 2D as 3D surface plot of full scan
  • 3D plot with Color and Density as done by true sampler

Training Path Visualization Module

• Load the Path
  • To store the regions searched in and plot optimization path
  • To map which combinatorial outputs were on the path and with what frequency
• Plot the Path
  • PCA of Parameters
  • Random axes projections of parameters
  • t-SNE of Parameters
• Map to the True answers
  • Use the combinatorial outputs and perform t-SNE on them over optimization path
  • (?) Plot path of subgraph features (key nodes/components), total cost, etc.

Handle the real world circumstances

• Noise Modelling
  • Swap out Parametrized Quantum Circuit with its Noisy version
  • Repeat all analysis above for noise aware environment (average for same state)