Code infrastructure and player algorithms for the Codenames board game.
This is the active fork of mkali-personal/codenames.
The solvers
module contains a multiple agent implementations, based on different strategies.
It is built above the codenames package, which contains the game logic and model definitions.
Given the board:
+-------------+------------+----------+--------------+----------------+
| ββ¬ money | βπ¦ drama | ββ¬ proof | βπ₯ baseball | βπ₯ imagination |
+-------------+------------+----------+--------------+----------------+
| βπ¦ steel | βπ₯ trail | ββ¬ giant | βπ¦ smell | ββ¬ peace |
+-------------+------------+----------+--------------+----------------+
| ββ¬ right | ββ¬ pure | βπ₯ loud | βπ afternoon | βπ₯ constant |
+-------------+------------+----------+--------------+----------------+
| βπ₯ fabric | ββ¬ violent | βπ₯ style | βπ¦ musical | βπ¦ commitment |
+-------------+------------+----------+--------------+----------------+
| βπ¦ teaching | βπ¦ africa | βπ¦ palm | βπ¦ series | βπ₯ bear |
+-------------+------------+----------+--------------+----------------+
A NaiveHinter
playing for the blue team will output "role", 4
.
From the logs:
Creating proposals for group size [4]...
Creating proposals for group size [3]...
Creating proposals for group size [2]...
Creating proposals for group size [1]...
Got 49 proposals.
Best 5 proposals:
('drama', 'musical', 'commitment', 'series') = ('role', 9.34)
('drama', 'musical', 'series') = ('films', 8.09)
('drama', 'musical', 'series') = ('comic', 8.04)
('drama', 'commitment', 'teaching') = ('focuses', 7.88)
('musical', 'commitment', 'teaching') = ('educational', 7.87)
Hinter: [role] 4 card(s)
Some extra data from the solver about the picked hint:
{
"word_group": ["drama", "musical", "commitment", "series"],
"hint_word": "role",
"hint_word_frequency": 0.999,
"distance_group": 0.194,
"distance_gray": 0.207,
"distance_opponent": 0.23,
"distance_black": 0.383,
"grade": 9.337,
"board_distances": {
"drama": 0.151,
"musical": 0.166,
"commitment": 0.189,
"series": 0.194,
"peace": 0.207,
...
"trail": 0.425,
"smell": 0.451,
"palm": 0.487
}
}
Find usage examples in the playground
directory.
Based on Google's word2vec embedding.
Hint generation:
- For each card subset
group
of size{4, 3, 2, 1}
from my unrevealed cards, collect hint proposal:- Find the mean of
group
's word embeddings. - Find the closest word to this mean (this will be the
hint
). - Calculate the distance from
hint
to all other unrevealed cards on the board. - Ensure the inspected
group
is the closest to the proposedhint
. - Ensure opponent cards, gray cards, and black card distance to
hint
are greater than a specified threshold. - Grade the
hint
proposal (based on the number of cards ingroup
and the distances fromhint
to the different word groups).
- Find the mean of
- After collecting all hint proposals:
- If no legal proposal was found, repeat step
1.
without filtering minimal distances (collect "dangerous" hints that might be confused with opponent cards). - Otherwise, pick the proposal with the highest grade.
- If no legal proposal was found, repeat step
Guess generation:
- Given a
hint
and number of cards, calculate the distance betweenhint
and all unrevealed cards on the board. - Iterate on number of cards:
- Pick the closest card to
hint
.
- Pick the closest card to
- Skip the extra guess.
Based on OpenAI's ChatGPT API. Doesn't work very well.
Clone this repository: git clone https://github.com/asaf-kali/codenames-solvers
.
- Make sure you have Poetry installed on your machine.
- Create a virtual environment (Python >= 3.8).
- Install dependencies using
make install
(or run the commands from theMakefile
). - Get a language model: TODO.
- Inside the
playground
directory, you will find different examples of how to use the solvers.
Currently, this project is not published on PyPI.
From your project virtual env, install the package with pip install -e <path_to_this_repo>
.
This needs to be updated.
- Download the zip file.
- Extract the
GoogleNews-vectors-negative300.bin
file into thelanguage_data/
folder and rename it toenglish.bin
.
Look in the GitHub repo.
Run in terminal: make video-render
.
If that didn't work: try python -m manim videos/explanation.py KalirmozExplanation -pql\h
.