Abstract:
An agent interacting with its surroundings to learn how to make decisions is called a reinforcement learner. Through reinforcement learning, the agent gradually gains the ability to behave in a way that maximizes its overall reward. Instead of receiving instructions from an instructor, it learns by trial-and-error process. VizDoom is an easy-to-use reinforcement learning-based research platform for Doom games. The task is to train agents in different VizDoom scenarios, including basic scenario, defend the center, and deadly corridor making decisions based on visual information from the screen. This project compares three common reinforcement learning algorithms. Double Dueling Deep Q Learning Network (DDDQN), Advantage Actor Critic (A2C), and Proximal Policy Optimization (PPO). We examine each algorithm's effectiveness in enhancing agent performance in the VizDoom environment through testing and analysis. Each of the previously stated algorithms is used to train the agents in our study, and their performance is assessed using a range of measures, including average score, number of kills, and reward. To further optimize each algorithm's performance, we have tried different values for tuning hyperparameters during our training process. During the training process, we changed critical hyperparameter values that influence the overall agent behavior and lead to its improvement. In the context of improving agent performance in VizDoom, the study's conclusions offer insightful information about the advantages and disadvantages of PPO, A2C, and DQ. Furthermore, our insights into hyperparameter optimization clarify the significance of parameter adjustment to optimize the performance of reinforcement learning algorithms in difficult real-world scenarios. The present study explores how reinforcement learning (RL), a type of AI training, can be better applied to video games. By understanding how RL works in games, researchers can develop better ways to train AI agents, making them smarter and more fun to play against.
