This code implements card-playing AI using modern AI/machine learning techniques. This was the solo final project for a graduate AI course. The basic implementation (built earlier in the course) for the card game player is in the rollouts-python repository.
The card game that the AI players play, is a monster-themed variation on the game Hearts aptly called Monster.
Download or clone the repository and install the packages listed in the 'requirements.txt' file. The Cython modules must be built using the command:
python3 setup.py build_ext --inplace
Then, the code can be run using the following command:
python3 main.py
This repository contains code for eight different card-playing AI. The baseline player, and perhaps the simpliest, is Grab And Duck. This is a simple greedy algorithm. It grabs points when it can, and avoids (or ducks) losses when it can. Another AI performs rollouts to make decisions, as often seen in backgammon opening strategy. Two AI players use a random-forests model trained on data generated from simulating games of Monster; the first uses the model alone, while the second pairs it with rollouts.
The implementation of Monster includes a number of improvements to boost performance,
extensibility, and readability. The game playing code and the AI
players have been implement using Cython; this package compiles Python
in C for a faster execution time on oft-reused code. One-thousand games of
Monster were played using only Grab and Duck players for a time of:
28401642 function calls (28388259 primitive calls) in 9.706 seconds
When the game playing code was setup to use Cython, the time was reduced to:
13198263 function calls (13184875 primitive calls) in 6.094 seconds
When the AI players were also setup to use Cython, the time was further
reduced to:
3168924 function calls (3155536 primitive calls) in 3.523 seconds
These times were collected using the cProfile utility included in Python.
Multiprocessing was also implemented, which provided a vast decrease in
the amount of time needed to run code. cProfile is more difficult to use for
multiple processes, so the following run has been recorded:
timothy@debian-desktop:~/Downloads/monte-carlo-python$ python3 main.py
(33.311600000000006, 736.158917532)
This images shows the results of 1,000,000 games of Monster, again with only Grab And Duck players. The win rate is, understandably, about 1/3. The execution time was about 12.5 minutes. Both Cython and multiprocessing allow for a small amount of compute resources (the computer used is seven or so years old) to go very far.