Artificial Intelligence and Machine Learning final project
Learning to Play 2048
Group members: Aashir Gajjar, Rohan Krishnan, Joel Feske
ABSTRACT
2048 is a puzzle game that has grown popular in the past few months. In this study, neural networks were used to learn strategies from human and AI players using both raw and processed data, and genetic algorithms were used to generate strategies from scratch. Games were simulated, and average performance was compared to an AI player which generated random moves. Neural networks were trained on varying numbers of games depending on the available data, and in general were able to successfully emulate the strategy they were trained on, with larger data sets producing more faithful emulation, as expected. Neural networks trained on processed data representing abstract qualities of the board out-performed networks trained only on board states. Genetic algorithms were also implemented to find novel strategies by attempting to predict the best next 6 moves based on the state of the current board. Both strategies learned by neural networks and strategies learned via the genetic algorithm were able to out-perform the random AI player.
2048 is a puzzle game that has grown popular in the past few months. In this study, neural networks were used to learn strategies from human and AI players using both raw and processed data, and genetic algorithms were used to generate strategies from scratch. Games were simulated, and average performance was compared to an AI player which generated random moves. Neural networks were trained on varying numbers of games depending on the available data, and in general were able to successfully emulate the strategy they were trained on, with larger data sets producing more faithful emulation, as expected. Neural networks trained on processed data representing abstract qualities of the board out-performed networks trained only on board states. Genetic algorithms were also implemented to find novel strategies by attempting to predict the best next 6 moves based on the state of the current board. Both strategies learned by neural networks and strategies learned via the genetic algorithm were able to out-perform the random AI player.
Your browser does not support viewing this document. Click here to download the document.
