Using Curriculum to Train Multisensory Foraging DRL Agents
Romulo F. Férrer Filho, Alexandre M. M. Santos, Halisson R. Rodrigues, Yuri Lenon Barbosa Nogueira, Creto Augusto Vidal, Joaquim Bento Cavalcante-Neto, Paulo Bruno S. Serafim · 2024
Deep reinforcement learning has shown great success in developing agents that can solve complex game tasks. However, most game agents use only visual sensors to gather information about the environment. More recent works have shown that agents that use audio sensors can perform better than vision-only agents. In this paper, we propose a curriculum-based training strategy to develop agents that effectively use audio as a source of information in foraging-based scenarios. First, we demonstrate that agents with both vision and hearing capabilities perform similarly to agents with only a visual sensor, indicating that the first ones ignore the audio. Then, we show that by using a gradually increasing difficult curriculum the agent effectively uses the audio information available, making it more robust to survive in scenarios where visual information is not available. Our results indicate that agents can be trained to effectively use audio as a source of information by using a curriculum based training strategy, improving their ability to deal with more tasks than agents with only vision.