Generation of Sokoban Stages using Recurrent Neural Networks

Abstract

Abstract-Puzzles and board games represent several important classes of AI problems, but also represent difficult complexity classes. In this paper, we propose a deep learning based alternative to train a neural network model to find solution states of the popular puzzle game Sokoban. The network trains against a classical solver that uses theorem proving as the oracle of valid and invalid games states, in a setup that is similar to the popular adversarial training framework. Using our approach, we have been able to verify the validity of a Sokoban puzzle up to an accuracy of 99% on the test set. We have also been able to train our network to generate the next possible state of the puzzle board up to an accuracy of 99% on the validation set. We hope that through this approach, a trained neural network will be able to replace human experts and classical rule-based AI in generating new instances and solutions for such games.

Procedural Level Generation for Sokoban via Deep Learning: An Experimental Study

“…There is a variety of generative models that rely on deep learning such as Variational Autoencoders (VAE) [4], Generative Adversarial Networks (GANs) [5], Auto-Regressive Models [14] and more. Many of these models has already been applied to PCG such as using GANs to generate Zelda levels [9], using VAEs for generating and blending levels from multiple games [15] and using LSTMs to generate Sokoban Levels [16]. All of the aforementioned methods rely on training the model to capture the distribution of the training data.…”

“…In this paper, we created an experimental setup to test some of the most recent deep-learning level generation methods but there are some techniques that we did not include such as LSTMs [16] and Adverserial Reinforcement Learning [31]. We are interested in adapting these methods to our experimental setting.…”

Procedural Level Generation for Sokoban via Deep Learning: An Experimental Study

“…There is a variety of generative models that rely on deep learning such as Variational Autoencoders (VAE) [4], Generative Adversarial Networks (GANs) [5], Auto-Regressive Models [14] and more. Many of these models has already been applied to PCG such as using GANs to generate Zelda levels [9], using VAEs for generating and blending levels from multiple games [15] and using LSTMs to generate Sokoban Levels [16]. All of the aforementioned methods rely on training the model to capture the distribution of the training data.…”

“…In this paper, we created an experimental setup to test some of the most recent deep-learning level generation methods but there are some techniques that we did not include such as LSTMs [16] and Adverserial Reinforcement Learning [31]. We are interested in adapting these methods to our experimental setting.…”