查看文章

The goal of this thesis was to research the possibilities of creating a reinforcement learning based agent, that could learn to play the 2048 board game. This game introduces a difficult task due to its factor of randomness. We designed and tested a base model of the agent using of a deep feedforward neural network, together with a few modifications of the original game and two types of input encoding in a number of experiments.

We implemented the agent using the Keras library with TensorFlow backend in Python language. For visualization of its performance we used Matplotlib library. We let the trained agents play 10000 games and compared their performance to an agent that selects actions randomly. Even as we did not achieve the expected level of performance with the original reward function, we were able to train the agent to achieve reasonably good results using its modification, which we consider an interesting finding. The best model reached the 2048 tile in more than 7% of the testing games.