Titolo della tesi: Tackling Sample Efficiency and Explainability in Reinforcement Learning through Integration with Planning
In modern robotics, the necessity for behaviors capable of generalizing and effectively managing complex situations has grown exponentially. Extensive research across various domains has addressed this challenge by exploring different decision-making techniques. While existing approaches to behavior generation and decision-making possess distinct strengths, they also exhibit certain drawbacks, such as computational complexity or the need for specialized problem-modeling expertise.
Automated Planning approaches heavily rely on environment modeling, whereas Reinforcement Learning methods eliminate this requirement but necessitate a large number of samples to learn effective policies. This thesis endeavors to leverage the strengths of both methods to overcome the weaknesses of each. The ultimate objective is to enhance the robot's learning speed for task resolution while obtaining important, comprehensible insights into the agent's behavior.
To assess the proposed methodologies, the research community advocates the utilization of diverse testbeds. In the context of this thesis, our focus lies on a range of demanding environments, with a notable emphasis on the RoboCup competition. This competition entails a fully autonomous five-on-five soccer game.
The game has progressively evolved into a more dynamic and active scenario within these Soccer Robots competitions. The diversity of challenges presented in RoboCup provides an excellent opportunity for testing and benchmarking the performance of complex decision-making algorithms. Participants can evaluate their algorithms' efficacy in multi-agent coordination, strategic positioning, and tactical decision-making.
Furthermore, the competitive aspect of RoboCup fosters innovation and drives the development of intelligent agents capable of sophisticated decision-making.