Codenames Bot

Start

python3 app/server.py

Algorithms

Clue-Giving

Clue-Giver ALGORITHM (don't mind if I do) Input the board Input which team to give clues for For each turn: Need to output one word + number of cards as a clue (absolutely -R) Receives guesses and marks the word as seen until no of valid guesses is elapsed chimken goes to the winner

The Magic Sauce ALGORITHM (perhaps lost, perhaps not)

Heuristic for the bot to rank its predictions
Bot needs to be able to evaluate an abitrary set of words up to a certain number against a set of clues

HIGH level approach of botto

Gather sets of cards to predict on
Make several predictions (clue) for each set of cards
And rank/choose the one most likely to give the highest success rate according to heuristic
Bot needs to try and account for negative constraints (avoid opposing colours, neutral, and meanie words)

GREEDY approach:

Create the superset of all card combinations
Generate N predictions for each set of card combinations indiscriminately
Rank predictions and select

Couple approaches:

Rank all possible sets of cards, pick the most similar set, then choose a clue for that set
Among set of cards, choose the top N most similar sets >1 to generate predictions

Try Word2Vec for generating clues https://stats.stackexchange.com/questions/267169/how-to-use-pre-trained-word2vec-model

Try Spacy https://spacy.io/

Classes of Approaches to Try

Aggregate clues by hand size, then normalize scores
Pick hands that are similar, then search for clues
Find a pure scoring method that drops the least similar subscore in hands of 3+
Let the user pick how many cards they want a clue for

Account For Danger Words

Goals:

Don't give clues that will lead to the other team's cards or the assassin card, preferably avoid neutrals as well
Pick the best clue that unifies as many of your team's clues as you can

Algorithm:

Store a map of clues to cards the clue is similar to:
    Good cards, danger cards, neutral cards
    For each card store a tuple of (card, card type, similarity rank, similarity score)
For each card that remains unflipped:
    For each of its top n similar words:
        Add an entry to the map for this clue, card, card type, similarity rank, and similarity score
Optionally: Filter out clues that apply to one or fewer good cards
Optionally: Filter out clues that apply to one or more danger cards
For each clue:
    Order the cards from most similar to clue to least similar to clue
    Calculate the number of consecutive good cards before hitting another type of card

Maybe we should also do another pass over clues and get the similarity of that clue to all cards? Then we can pick the clues that divide the cards into a ranking where the first, consecutive cards in the ranking are our team and then there is a large drop in similarity before we get to a danger or neutral card.

Scripts

Set Up Repository

# Set up virtual environment
python3 -m venv venv
source venv/bin/activate
pip3 install -r requirements.txt
# Download word embeddings
mkdir embeddings
wget -O embeddings/glove.6B.zip http://nlp.stanford.edu/data/glove.6B.zip
unzip embeddings/glove.6B.zip -d embeddings
rm embeddings/glove.6B.zip
# Make model from word embeddings
mkdir models
python3 make_model.py
# Run the main script
python3 main.py
# Deactivate virtual environment
deactivate

Get Cards From Codenames

// Open https://horsepaste.com
// Go to a game
// Click on Spymaster view
// Run this in the browser console:
Array.from(document.querySelectorAll('.word')).map(e => e.ariaLabel).join('\n')

Convert Embeddings to KeyedVector Model

glove_file = "/Users/vineshkannan/Documents/GitHub/codenames/glove.6B/glove.6B.100d.txt"
model = KeyedVectors.load_word2vec_format(glove_file, binary=False, no_header=True)

Make Environment Variables File

PORT=5000
ALLOW=http://localhost:5000,http://104.236.21.173:5000
CHROME=/Users/vineshkannan/Documents/dev/chromedriver

Download Chromium Webdriver

# https://chromedriver.chromium.org/downloads
mkdir driver
wget -O driver/chromedriver.zip https://chromedriver.storage.googleapis.com/100.0.4896.60/chromedriver_linux64.zip
unzip driver/chromedriver.zip -d driver
rm driver/chromedriver.zip
echo -e "CHROME=$(pwd)/driver/chromedriver" >> .env
# If Chrome (in addition to Chrome webdriver) is not installed:
# https://gist.github.com/ziadoz/3e8ab7e944d02fe872c3454d17af31a5

Get Random Words

// https://randomwordgenerator.com/
// Pick nouns, adjectives, or verbs, set word size by syllables or letters
Array.from(document.querySelectorAll(".support")).map(e => e.innerText)

Notes

Stop words list from sebleier
Emoji as favicon from Lea Verou, Chris Coyier

vingkan / goatnames