Autonomous Foraging with SARSA-based Deep Reinforcement Learning

• Anderson Mesquita UFC
• Yuri Nogueira UFC
• Creto Vidal UFC
• Joaquim Cavalcante-Neto UFC
• Paulo Serafim Instituto Atlântico

Resumo

This work proposes a system capable of autonomous behavior using Computer Vision and Deep Reinforcement Learning. These learned behaviors are those related to foraging in an environment that has food and poisons distributed throughout a scenario. We use Deep Learning framework to process color images. These images simulate the agents vision. The foraging task is modeled as a reinforcement learning problem, in which an input constituted by raw pixels is processed by a convolutional neural network resulting in a set of actions. A Deep Learning algorithm based on SARSA was used. During training, the agent selects the actions based on a probability distribution called Softmax. The objective of this work is to present an agent capable of searching for food and distinguishing it from poisons through continuous learning and without the help or external intervention from humans. The experiments show that the agent is able to distinguish food from poisons without the hints or markings in its vision input. This highlights the advantages of combining Deep Learning with reinforcement learning for the foraging problem. The results of this work form an initial basis for understanding the relationship among autonomy, cognition and perception in artificial agents.

Palavras-chave: autonomy, deep learning, reinforcement learning, computer vision