Generalized game-testing using reinforcement learning

Abstract

The gaming industry has experienced significant growth and evolution, becoming a prominent sector in entertainment and technology. This growth has led to increased consumer expectations regarding the quality and complexity of games, prompting developers to explore innovative solutions to meet these demands. To meet these demands, one of the pivotal approaches adopted by game developers is the game testing process. Game testing is an incredibly resource-intensive procedure, demanding comprehensive evaluation of all aspects of a game through actual gameplay. To address this challenge and alleviate the associated workload, this thesis proposes an innovative approach to game testing. This method integrates a generic environment framework with reinforcement learning (RL) models, facilitating seamless communication between any game and an RL model under specific conditions. The framework optimizes the game testing process by capitalizing on the efforts of game developers. It relies on developers to compile and transmit essential information, such as state and reward data, to the generic environment. Subsequently, this data is processed and harnessed within the RL model, allowing it to learn and play the game in accordance with developers' intentions, while simultaneously generating valuable data for game testing purposes. This method also capitalizes on the beneficial aspect of game-playing AI agents trying out various actions in different states as they learn to play games. Game testing entails the creation of diverse scenarios by implementing different actions in various in-game situations. These scenarios are observed, and, when necessary, actions are taken in the game development process based on these observations. Therefore, as the situation where game-playing agents experience various scenarios closely resembles game testing, we can utilize not only the actions performed by agents during testing but also their behaviors during training as part of the game-testing content. The experimental phase of the study involved the deployment of six distinct builds of the same game, each serving as a means to test the functionalities of the generic environment and observe their impact on the behavioral patterns of RL models. These builds were thoughtfully crafted to uncover various aspects of RL model behavior and the diverse methods of representing game states. These builds can be summarized as follows: - Basic side-scroller: This build's purpose is to test the seamless communication between the generic environment framework, the game build, and the RL model. It features a simple reward system designed to guide the player to a target point, an action space consisting of three actions, and employs a state image as the state information. Exploration-oriented side-scroller:} Designed to encourage the player to explore the entire game area, this build incorporates a comprehensive reward system. It boasts an action space comprising four actions and utilizes a state image as the state information. - Exploration-oriented side-scroller with colored textures: This build serves as a variant of the exploration-oriented side-scroller build, with the only alteration being the modification of game textures. Its purpose is to investigate the impact of texture changes on the training of RL models. - Goal-oriented side-scroller: Sharing the same action space and state information as the exploration-oriented side-scroller build, this build primarily aims to observe the effects of reward system modifications. It employs a detailed reward system to guide the player toward specific objectives and a goal. - Exploration-oriented side-scroller using no image: With an identical action space and reward system structure as the exploration-oriented side-scroller build, this build seeks to examine how using a state array as state information influences the RL model's behavior. - Exploration-oriented side-scroller using image and array: Being similar to the exploration-oriented side-scroller build in action space and reward system structure, this build aims to maximize its impact on the RL model's behavior. It achieves this by employing both a detailed state array and a state image as state information. - Arcade: This build aims to demonstrate how the generic environment framework will perform in a completely different game. It has both exploratory and goal-oriented structures. It features a moderately complex reward system and an action space consisting of five actions. It uses both arrays and images as state information. The investigation into the communication system between the RL agent and the game build yielded valuable insights. It became evident that the generic environment framework played a crucial role in achieving positive and efficient outcomes. Nevertheless, the research also pinpointed areas ripe for enhancement, particularly concerning the reduction of the workload on game developers and the resolution of issues stemming from external factors. The logging system integrated into the generic environment has proven to be a valuable asset in the realm of game testing. It leverages the total reward accrued in each episode, efficiently guiding the selection of episodes meriting closer scrutiny. Furthermore, the supplementary information provided by this system offers exceptionally insightful data, greatly enhancing our comprehension of the actions taken in various gaming scenarios. Our proposed approach holds significant potential in the realm of game testing. It enables AI agents to adjust their behaviors by utilizing dynamic rewards and extensive state information from arrays and images to meet specific criteria. Moreover, successful game-testing outcomes have been consistently observed throughout both the training and testing phases, where agents adeptly exploit game vulnerabilities and uncover unforeseen features and bugs. In spite of the apparent successful outcomes, the implementation involving both a state image and a state array exhibited a notable reduction in training speed and encountered a substantial level of system load attributed to hardware constraints during the training process. When evaluated in accordance with the objectives of the thesis, it can be concluded that, overall, the proposed method has achieved successful outcomes in the game testing process and holds promise for future development potential. Further endeavors aimed at enhancing system performance may yield positive results concerning the broader applicability of game testing.