Treinando agentes de forrageamento multissensoriais usando currículo

Abstract

Deep Reinforcement Learning (DRL) has demonstrated significant success in the development of agents capable of solving complex game tasks. However, the majority of game agents rely solely on visual sensors to gather information about the environment. Recent studies have indicated that agents utilizing audio sensors can outperform those that depend exclusively on vision. This work evaluates the ability to utilize multiple sensors in a foraging task. It also proposes a curriculum-based training strategy to develop agents that effectively use audio as a source of information in foraging scenarios. First, it is demonstrated that agents with both visual and auditory capabilities perform similarly to agents with only visual sensors, indicating that the former ignores the audio input. Initially, it is demonstrated that agents equipped with both vision and hearing capabilities perform similarly to agents with only visual sensors, suggesting that the former tend to ignore the audio input. Subsequently, it is shown that by implementing a gradually increasing difficulty curriculum, the agent effectively utilizes the available audio information, enhancing its robustness in scenarios where visual information is absent. The results indicate that agents can be trained to effectively use audio as a source of information through a curriculum-based training strategy, thereby improving their capacity to handle a wider range of tasks compared to agents with only vision.