In this thesis, the application of the reinforcement learning agent to the Sokoban game was investigated. The paper describes the theory, basic principles and application techniques of reinforcement learning. Also, different neural networks are described, due to their use in the architecture of reinforcement learning strategies. Thesis reviews Sokoban game environments and state-of-art reinforcement learning libraries, frameworks. The paper compares implementations of different reinforcement learning algorithms and different strategies in the Sokoban game environment. The obtained results of transfer learning are compared to those of an agent trained in an environment of equal complexity without transfer learning. Analysis of the results showed that using transfer learning for the Sokoban game environment, reinforcement learning algorithms can achieve better results.


    Access

    Download


    Export, share and cite



    Title :

    Skatinamojo mokymosi algoritmų skirtų Sokoban žaidimo agento valdymui palyginimas ; Comparison of reinforcement learning algorithms for sokoban game agent training



    Publication date :

    2020-06-10


    Type of media :

    Theses


    Type of material :

    Electronic Resource


    Language :

    Lithuanian , English


    Classification :

    DDC:    629