Codenames AI

An interactive command-line game of codenames, with an AI implementation that can play as a code giver or guesser.

You can see it in action at http://personal.mshafir.com/codenames/ though my AWS instance is a tiny so it'll be slow and might run out of memory.

Requirements

Vectors

You need to download a set of word vectors in order to run the code:

• You can download the GloVe vectors here: https://nlp.stanford.edu/projects/glove/

  • the 6B token vectors work okay
  • the 42B work better
  • I imagine the 812B work best, but haven't verrified that

• LexVec is another option, can be downloaded here: https://github.com/alexandres/lexvec

Words

This repo comes with a sample word list

Using a larger word set, like the ones from https://github.com/dwyl/english-words will also improve the quality of the clues

Word Counts

Word counts help give the code giver an idea of how common or esoteric different words are, it's configured to work against the following count list:

http://norvig.com/ngrams/count_1w.txt

Conda

1. get miniconda: https://conda.io/miniconda.html
2. conda env create -f environment.yml --force

Execution

Pre-compile distance matrix

You will need to compile a distance matrix to run things. This is a one-time cost of around 30 minutes for large vectors.

python compile_distances.py

Cut distance matrix

You can speed up loading and execution by getting rid of matrix entries above a certain distance.

First argument is the distances file, second is the cutoff.

python cut_distances.py resources/cached.dist 0.9

Run

You can update main.py to use the target distances and word counts file you want. Set debug=True on a codegiver AI to see the scores and what it's thinking.

Update codegiver.py to change the thresholds for clues

python main.py