Search

Rad se fokusira na korištenje genetskog algoritma za treniranje neuronske mreže za igranje popularne igre 2048. Genetski algoritam koristi prirodnu selekciju, križanje i mutaciju kako bi generirao nove generacije mreža koje su sve bolje u igranju igre. Kroz iterativni proces, mreže uče strategije i pravila igre koje su potrebne za postizanje visokih bodova. Nakon što je mreža trenirana, testira se njezina sposobnost igranja igre i ocjenjuje se njezin uspjeh u usporedbi s drugim metodama igranja. Rezultati projekta pokazuju da genetski algoritam može biti uspješan u treniranju neuronske mreže za igranje 2048 igre. The project focuses on using a genetic algorithm to train a neural network to play the popular game 2048. The genetic algorithm employs natural selection, crossover, and mutation to generate new generations of networks that get progressively better at playing the game. Through an iterative process, the networks learn strategies and rules of the game that are necessary to achieve high scores. Once the network is trained, its ability to play the game is tested, and its performance is evaluated in comparison to other playing methods. The project's results demonstrate that the genetic algorithm can be successful in training a neural network to play the 2048 game.

Este proyecto ha tenido como objetivo principal la construcción de un robot capaz de jugar autónomamente al conocido juego 2048. Se puede desglosar el trabajo en dos grandes partes: una de software y una de hardware. En lo que respecta a la primera, se ha realizado un análisis inicial de diversos aspectos del juego y se ha comparado distintos métodos de resolución, principalmente minimax, expectimax y redes neuronales. De todos estos métodos se han propuesto multitud de variantes en las que se han añadido diversas características como, por ejemplo, las distintas heurísticas para evaluar los nodos del árbol de búsqueda en los métodos minimax y expectimax o distintas arquitecturas en el caso de las redes neuronales. En cuanto al hardware, se ha construido un robot ideado para ser posicionado delante de la pantalla de un ordenador con el juego abierto e interaccionar con el ordenador, dando las órdenes de movimiento a través del puerto USB. El robot incorpora una placa Raspberry Pi 4, una placa Pro Micro ATmega32u4, una cámara y componentes de soporte, todo ello, dentro de una carcasa hecha a partir de modelado e impresión 3D. The main objective of this project was the construction of a robot capable of autonomously playing the well-known game 2048. The work has two main parts: software and hardware. Regarding the first, an initial analysis of various aspects of the game has been carried out and different resolution methods have been compared, mainly minimax, expectimax and neural networks. Many variants of all these methods have been proposed in which various features have been added, such as, for example, the different heuristics to evaluate the nodes of the search tree in the minimax and expectimax methods or different architectures in the case of neural networks. For the hardware, a robot has been built and designed to be positioned in front of a computer screen with the game open and interact with the computer, giving movement orders through the USB port. The robot incorporates a Raspberry Pi 4 board, a Pro Micro ATmega32u4 board, a camera and supporting components, all inside a 3D modelled and printed case.