gullesnuffs/codenames

A codenames bot playing the part of the spymaster.

codenames's Issues

Better makefile

Should really do dependency tracking, and build .cpp files into .o files.

Currently we build each file twice, due to calc.cpp; builing .o files would fix that as well.

Normlize norms on board words

Some words on the board are quite vague and far from other words. Can we increase their norms, to give equal consideration to all words on the board?

Implement MCTS

Use hinted_words on web client

Instead of trying to maximize the number of words in one turn, try to plan ahead and minimize the number of turns until victory.

Don't sacrifice performance in inappropriate engine

It's not appropriate. (Maybe make a Dictionary class?)

"No more clues"

https://judge.omogenheap.se/codenames/api/1/clue?engine=glove&color=r&colors=ccccbrar&words=circle%2Ctorch%2Ceagle%2Cflute%2Cbeijing%2Civory%2Corange%2Cundertaker&hinted_words=&index=1&count=1

It finds a clue for index=0, but fails for index >= 1.

Use a thesaurus for words on the board

To more explicitly test all possible meanings of them. When you play codenames, you try to give more equal consideration to obscure meanings of words and to common ones than you'd normally do.

Reduce syntactical nature of clues

E.g. verbs are currently similar to other verbs, which is natural, but not what we prefer for codenames clues. Maybe one can do some smart linear algebra thing??

Compile error on MacOS

I get the following error when trying to use the default g++ on my up-to-date version of MacOS.

codenames.cpp:1:10: fatal error: 'bits/stdc++.h' file not found
#include <bits/stdc++.h>

Manual training

Let's say we do a couple of hundred manual rankings of orders for the top 5-10 words for various clues. We could then:

• use this for parameter tweaking: try to maximize some sort of resulting score
• maybe use it for finding interesting and uninteresting relations in the underlying vector space

Aron suggests using Mechanical Turk, not so much because it gives good results as because it would be really fun.

Multi-threading

Don't use data.bin

We should take the model as a command line argument instead (for both ./calc and ./codenames)

Useless clues in conceptnet

Seems like it ought to weigh more specific words higher, e.g. like how glove uses norms. Or compare nearness with a useless string?

Investigate glove norms

"gotta" is a terrible clue. Maybe it has low norm also, because it doesn't mean very much? Let's make the program print the norms and see if we can make out anything from them.

Lower-case words in glove

For usability we want only lower-case words. But glove contains both upper- and lower-case words, and sometimes with different meaning... Take weighted averages, or just the most common version?

Desperation mode

When the opponent has only one or two clues left:

• give gray, opponent and assassin the same penalties
• be more optimistic, and try to give a clue that covers everything not already covered
• maybe even try to give a clue for the complement

Minimize disk usage

So it fits on a phone, for instance. Currently the word2vec model uses ~50MB, which we could reduce by:

• using fewer words, winning say a factor of 2
• using half-floats instead of floats, winning a factor of 2
• storing similarities between each dictionary word and Codenames word instead of vectors, winning a factor of 300/100=3 (makes it impossible to play with custom words, though)

Use several models at once

Use list of inappropriate words to avoid using such words as clues.