Easy Learning of Reinforcement Learning with a Gamified Tool

• Elson Almeida Dreveck Unilasalle
• Alex V. Salgado Unilasalle / UFF
• Esteban W. Gonzales Clua Unilasalle
• Luiz Marcos Garcia Gonçalves UFRN

Resumo

The use of the AWS environment with DeepRacer provides an interesting educational learning platform, which allows to train and apply machine learning models in vehicles that compete in a virtual race. In this work we show the stages of creation and training of reinforcement learning models through this platform and validate its use through a qualitative evaluation. A user without prior knowledge in reinforcement learning obtains reasonable performance results on the competitions using the PPO and SAC algorithms. The tool allows to add entropy adjustments, besides the definition of an state-action space with associated reward functions. To this end, these preliminary qualitative results support and motivate the use of this tool that makes easy the learning of reinforcement learning being thus of great interest and importance to the Robotics community.

Palavras-chave: Training, Conferences, Reinforcement learning, Tools, Entropy, Performance analysis, Robots