experiments

Dominoes ML Experiments

This folder contains python scripts for running experiments with the dominoes package. Here, you can find a brief table of contents that explains the purpose of each experiment.

Standard usage

To run these experiments, you should already have cloned the dominoes repository and created an environment to run it (see instructions in the top level README). Next, navigate to the wherever your repository is cloned (e.g. where you can see the dominoes, docs, & experiments folders).

Finally, you can run each of these experiments with the following command, where <experiment_name> is replaced with the name of the file, and optional arguments are defined in the ArgumentParser for each file.

python experiments/<experiment_name> --optional --arguments

Experiments

This is a very brief table of contents explaining the goal of each experiment. The point of this section is just to give a rough idea for what each file does. More detail on the results is explained in the documentation files (you can find these files listed in the "Documentation" section of the repositories README file). And of course, you can look through the files to see what their code does (I think they're pretty well commented...). If you have a question about an experiment or an idea for a new one, let me know!

Constructing a Dominoes RL Agent

• Basic Agent ELOs
• Training a TD-Lambda Agent
• Performance of TD-Lambda Agents

Pointer Network Analysis

• Pointer Network Demonstration
• Architecture Comparisons
  • With Reinforcement Learning
  • With RL and Uneven Batch Size
  • With Supervised Learning

Additional Curiosities

• Best Line Agent ELOs
• Table Position Matters