An evaluation of deep reinforcement learning & neuroevolution in stealth game problems

• Autor(a) principal: Silva, Pedro Miguel Figueiredo da
• Data de Publicação: 2024
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Texto Completo: http://hdl.handle.net/10400.26/52330
• topic: Artificial neural networks; Deep learning; Deep reinforcement learning; Neuroevolution; Stealth games

Deep learning (DL) is a rapidly evolving field with the potential to revolutionize the videogame industry. While there have been efforts by game companies to research DL approaches to game implementation, very little has been used in the industry. One of the main reasons is the lack of diversity in the game genres to which these algorithms are tested and applied to. Stealth games are one of the most popular gaming genres, characterized by the need to balance both states of sneaking and detection. Within this genre, game designers have traditionally relied on quality assurance testers to understand how players interact with stealth levels, which can be a time-consuming process. Potentially DL could be used to facilitate the game development processes and make them more cost-effective. This project aims to investigate whether DL can be used in game development, particularly in stealth games. As such, the project will focus on developing diverse stealth game environments, optimizing DL algorithms for these environments, comparing different DL algorithms, and measuring the computational cost of these algorithms. To achieve the stated objectives, the project follows a structured methodology: developing diverse stealth game scenarios, implementing state-of-the-art DL models, conducting hyperparameter tuning, performing performance evaluation, and conducting computational benchmarking. The results obtained from following this methodology show that in most cases, the implemented DL algorithms were able to learn all the different stealth game levels, indicating the strength and flexibility of the implemented algorithms. Also, a deeper evaluation of the algorithm's performance showed clear distinctions into which DL algorithms provided better performances, namely the Rainbow-DQN algorithm within the deep reinforcement learning and GA algorithm for neuroevolution. Regardless of the successful results, further research is needed to better comprehend the possibilities that the DL techniques can offer to game development in the context of stealth games.