Benchmarking reinforcement learning algorithms in first-person shooter games using VizDoom

Computer games are considered one of the best test beds for evaluating artificial intelligence algorithms, as it is a well-known practice before applying the algorithms in the real world, such as the robotics industry. A machine learning technique, known as reinforcement learning, utilizes positive and negative rewards to guide an artificial intelligence agent as it learns new tactics and strategies. This study compares four reinforcement learning algorithms: Dueling Double Deep Q-Network (Dueling DDQN), Advantage Actor-Critic (A2C), LSTM-Based Advantage Actor-Critic (A2C LSTM), and REINFORCE. The game artificial intelligence (Game AI) based platform VizDoom evaluates and compares these reinforcement learning algorithms. VizDoom is based on the first-person shooter (FPS) video game Doom, which has had a significant influence on artificial intelligence. The results are compared, and, in most cases, Dueling DDQN outperformed all other algorithms in all chosen scenarios. However, in contrast, the A2C performed well for the kills metric in the defending the center scenario only. Finally, the proposed work’s analysis, implications, and limitations are presented, along with the potential future directions for research.