Assessment of Various Deep Reinforcement Learning Techniques in Complex Virtual Search-and-Retrieve Environments Compared to Human Performance

Recently, the area of deep reinforcement learning (DRL) has seen remark-able advances in fields like medicine, robotics, and automation. Nonetheless, there remains a dearth of understanding about how cutting-edge DRL algorithms, such as Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC), match up against human proficiency in demanding search-and-retrieve operations. Furthermore, there is a scarcity of structured assessments of the efficacy of these algorithms in intricate and dynamic surroundings after hyperparameter adjustment. To tackle this discrep-ancy, the research at hand aims to evaluate and contrast the impact of the proportion of targets to distractions on human and machine agents’ performance in a complex search simulation using a professional gaming engine. Moreover, the influence of the amount of neurons and layers in DRL algorithms will be scrutinized in connection to the search-and-retrieve missions. The task requires an agent (whether human or model) to traverse an environment and gather target objects while sidestepping distractor objects. The results of the study exhibit that humans accomplished better in training scenarios, whereas model agents performed better in test scenarios. More-over, SAC emerged as a superior performer compared to PPO across all test conditions. Furthermore, boosting the amount of units and layers was found to enhance the performance of DRL algorithms. These conclusions imply that similar hyperpa-rameter configurations can be utilized when contrasting models are generated using DRL algorithms. The study also delves into the implications of utilizing AI models to direct human decisions.