Project name: Paper football Authors: Maciej Dobrzański, Izabela Dąbrowska

Overview

The result of our project is a program to play the game called paper football with choosen computer opponent or to simulate a match between two choosen bots. The rules are described here: https://en.wikipedia.org/wiki/Paper_soccer. Additionally we included framework allowing us to create and train our own bots based i.a on methods of artificial intelligence. The result are also a few bots which we have already trained.

Project structure

We divided our project into model, view and controller packages. Below, there is a more detailed description of files included in the project.

File path Description

view* Folder with files for GUI model\game.py Interface for files: model\gameState.py and model\gameStateUtils; Includes functions to get allowable moves (on 3 modes: simple, deepcheck and random), to make moves etc. model\ml-module\agent.py Class standing for computer player. There are 3 modes of getting a move and2 modesof choosing one (described below). model\ml_module* Implementation of predictors, MCTS, memory, deterministic models controller\playing.py Set of functions for playing training matches (with saving to memory and retraining) and validation ones controller\main_training.py File with training loop for models controller\main_play.py Start of a game controller\config Includes configuration options for the first game (depth of search, max times, learning parameters

There are 3 modes of getting allowable moves avaliable for user and 2 modes of choosing one of them. Getting allowable moves ‘simple’ - all allowable moves that are not losing. If there is the winning move, it should be returned as the only one. ‘deepcheck’ - all allowable moves that are not losing and the moves and the moves after which the opponent has not winning move. ‘random’

Choosing a move

‘model’ - we choose move with the best result returned by the model (‘probability of win’) ‘mcts.simple’ - simple MCTS ‘mcts.boosted’ - MCTS, where the initial weights are initialized from the model

Models

We trained two model, one dedicated for simple model evaluation mode of the agent, and secound for mcts boosted approach.