SyntheticPreferenceGenerator

An extensible python program which generates sets of synthetic preference examples from randomly generated model of specified preference languages.

Dependencies:

Pytorch: https://pytorch.org/

GenCPNet: https://github.com/nmattei/GenCPnet

Torch-Sparse: https://github.com/rusty1s/pytorch_sparse

Torch-Scatter: https://github.com/rusty1s/pytorch_scatter

Torch-Geometric: https://github.com/rusty1s/pytorch_geometric

Usage:

python3 SynthPrefGen.py [options] [config file]

• Options: -p problem subproblem specifies the problem and subproblem to run.

  Problem Numbers:

  Problem Number	Description
  1	Learning Preferences with Neural Networks.
  2	Learning Preferences with Lexicographic Preference Models.
  3	Learning Preferences with Simulated Annealing.
  4	Various related baselines.

  Subproblem numbers for problems 1-3

  Subproblem Number	Description
  1	Learning Preferences from a single agent.
  2	Learning Joint Preferences (Utilitarian).
  3	Learning Joint Preferences (Maximin).
  4	Learning Preferences from a single agent (Full Evaluation).

  Note: Subproblem 3 for problem 1 is not implemented.

  Subproblem for problem 4:

  Subproblem	Description
  1	Builds and analyzes full fitness graph.
  2	Builds and analyzes partial fitness graph using hill climbing.
  3	Analyzes basic hill climbing and random restart approach.
  4	Example set classification using custom feature vectors.
  5	Example set classification using custom feature vectors (complete relation example set, for learning).
  6	Example set classification using graph convolutions.
  7	Example set classification using graph convolutions (complete relation example set, for learning).

  -i [config file]

  Specifies the type of model to be learned by simulated annealing (default: 113RPF_learn.config). Note: Only applies to problems 2 and 4.

  -l [n]

  Specifies the number of hidden layers to add to the neural network (default: 3, max: 3). Note: Only applies to problem 1.

  -o [filename]

  Specifies the name of the data output file (default: a.out).

#Warning This software is provided as-is and while efforts have been made to make it more user friendly, improper usage may render undesired results.