Artificial Intelligence and Machine Learning - Project

Dice Wars

Dice Wars is a strategy game where players take turns to attack adjacent territories to expand their area. Each territory contains a number of dice determining player's presence and strength. The objective of the game is to conquer all territories and thus eliminate each opponent.

This is a client-server implementation that based on a bachelor's thesis at FIT BUT.

Installation

To use this, you need to have python3 and the following python packages:

hexutil
numpy
pyqt
matplotlib

A standard requirements.txt is provided.

Furthermore, the root of the repository needs to be in PYTHONPATH.

As an easy way of setting up the environment, do the following:

# install
mkdir SUI
cd SUI
git clone https://github.com/ibenes/dicewars.git repo
cd repo
bash install.sh

# setup
. path.sh
# try a game
python3 ./scripts/dicewars-human.py --ai dt.sdc dt.rand xlogin00 xlogin42

Running the game

There are three different scripts prepared, which allow for testing different scenarios. However, they all expose a common set of parameters for controlling the pseudo-randomness in the game:

-b  geometry of the board
-o  clustering of areas into possesion of individual players  
-s  assignment of dice to areas

When not set, the source of pseudo-random numbers is seeded from current time, becoming effectively random.

Finally, individual AIs are refered to as follows: For every module in dicewars.ai, which contains a class AI, the AI is identified by module. Examples are given throughout the following sections.

Playing with human

Starts a human-controlled client along those driven by AIs. There can be between 1 and 7 AIs. For an easy game (beware, defeat is still a real possibility), try:

python3 ./scripts/dicewars-human.py --ai dt.sdc dt.rand xlogin00 xlogin42

For a really challenging setup, try:

python3 ./scripts/dicewars-human.py --ai dt.sdc dt.ste dt.stei dt.wpm_c

Playing with fixed AI order

Starts a set of games between AIs in given order. Increments the board seed with every game. Additionally exposes these options:

-n      number of board to be played
-l      folder where to put logs of last game
-r      keep reporting which game is being played

An example:

python3 ./scripts/dicewars-ai-only.py -r -b 11 -o 22 -s 33 -c 44 -n 10 -l ../logs --ai dt.stei xlogin42

Running a tournament

Keeps picking a subset of AIs of specified size and has them play together. The total set of AIs considered is given in the script itself. Ownership (-o), dice assignment (-s), and dice rolls (-f) seeds are not exposed. Additionally exposes these options:

-n      number of boards to be played
-g      size of games in number of players
-l      folder where to put logs of last game
-s      seed for selecting who plays whom
-r      keep reporting what game is being played
--save  where to save the resulting list of games

For every board, all rotations of a random permutation of the player order are played, thus the total number of games equals N x G

An example:

python3 ./scripts/dicewars-tournament.py -r -g 2 -n 50 -b 101 -s 1337 -l ../logs --save ../tournaments/tournament-g2-n5000.pickle

This script can also be used for evaluation of a specific AI, ensuring that it takes part in every game played. This is achieved through --ai-under-test, e.g.:

python3 ./scripts/dicewars-tournament.py -r -g 2 -n 50 --ai-under-test dt.sdc -b 101 -s 1337 -l ../logs

Observing convergence of winrates

If you have saved games from a tournament (through its --save option), you can display the evolution of the winrates:

python3 ./scripts/winrate-progress.py --xmin 10 ../tournaments/tournament-g2-n5000.pickle

Note that the evolution of winrates does not have any other interpretation than the rate of convergence!

Implementing AIs

See dicewars/ai/template.py and other existing AIs in the package. An AI is a class implementing two standard functions: __init__() and ai_turn()

Name vs. instance

Players and areas exist primary as instances of Player and Area. However -- originally for serialization purposes -- they are both referred to by their "name". These names are instances of int. Board can return Areas as given by name` every Area knows its name.

There is no reason for an AI to access instances of Player.

AI interface

The constructor is expected to take following parameters:

player_name     the name of the player this AI will control
board           an instance of ``dicewars.client.game.Board``
players_order   in what order do players take turns

The turn making method is expected to take following parameters:

board               an instance of ``dicewars.client.game.Board``   
nb_moves_this_turn  number of attacks made in this turn
nb_turns_this_game  number of turns ended so far
previous_time_left  time (in seconds) left after last decision making

The AI.ai_turn() is required to return an instance of BattleCommand or EndTurnCommand.

Multi-module implementation is possible, see xlogin42 for an example.

Learning about the world

Board's get_player_areas(), get_player_border(), and get_players_regions() can be used to discover areas belonging to any player in the game. Instances of Area then allow inquiry through get_adjacent_areas(), get_owner_name() and get_dice().

It may also be practical to acquire all possible moves from dicewars.ai.utils.possible_attacks(). This module also provides formulas for probability of conquering and holding an Area.

The instance of Board passed to AI is a deepcopy, so the AI is free to mangle it in any way it deemed useful.

Debuging visually

In addition to whatever favourite debugging method you have, Dicewars provide a simplistic way of visually inspecting the state of the game. There is save_state() function provided by dicewars.ai.utils, which creates a dump of the state which AI observes.

The saved state can than be loaded by scripts/visual-debugger.py. This visual debugger allows displaying different information on areas (change label through the sole button in the interface) and a custom detailed information upon selection of an area of interest.

There is an example of saving games in dicewars.ai.xlogin42.phased.AI, and the visual debugger displays a bit of information this AI cares about. It is expected that the developers of new AI will adjust the debugger's DetailedAreaReporter to the needs of their AI.

Dealing with misbehaving AIs

Slow AI -- AIs have a fixed 10s time for constructing themselves. Additionally, a Fischer clock of 10s and 0.1s increment limits time for decision making. AI failing to make a decision will be stopped in deciding and a EndTurnCommand will be sent instead (but the increment is made anyway, so the AI will be able to continue playing). AIs are informed about the time they have left through time_left.
Stupid AI -- AI attempting to make an illegal move will be switched off, idling for the rest of the game.
Passive AI -- AI sending only EndTurnCommand will be quickly taken care of by other players. However, if no AI makes a move for 8 consecutive rounds, the game will be contumated and every player scores a defeat.

xstupi00/Dice-Wars