Transferring the knowledge acquired by a robot in previous experiences to new tasks is one of the challenges faced by the Artificial Intelligence field nowadays. This scientific research project has the objective to propose, develop and evaluate abstraction methods applied to the transfer of policies learned in some tasks by the agent to new, similar and unseen tasks, using the Reinforcement Learning framework. Regarding the abstraction problem, first-order logic will be particularly used as the representation language, in order to achieve generalization of tasks through the power of abstraction of the representation. This knowledge transfer will be explored in various domains, especially the robot navigation domain in places such as buildings and hospitals.
News published in Agência FAPESP Newsletter about the scholarship: