Working towards General AI through Deep RL 🤖

Recommender Systems are widely used to filter information and suggest personalized content to users. Traditionally, these systems focus on maximizing relevance for individual users, which can create popularity bias and lack of diversity in recommendations. This work proposes a recommendation framework, A2Fair, that seeks to balance recommendation relevance with fairness in the exposure of recommended items. The approach consists of an Actor-Critic Reinforcement Learning algorithm, where state representation incorporates features of the last items interacted by the user, as well as the current fairness situation among item groups. The action suggests the next item to recommend, while the adaptive reward function weighs relevance and fairness according to the user's affinity for diversity. The A2Fair framework enables personalized and adaptive recommendations, taking into account individual preferences while promoting fairness and diversity. Experiments conducted on public datasets demonstrated the effectiveness of the A2Fair framework compared to other methods.

In large-scale problems, reinforcement learning systems must use parameterized function approximators, such as neural networks, to generalize across similar situations and actions. Despite proving to be an effective method in specific complex problems, they still fail in generalization, and the most common reinforcement learning evaluation environments still encourage training and evaluation on the same set of environments. To assess an algorithm's ability to generalize across tasks, evaluation environments that measure performance on a distinct test set from those used in training are necessary. This work evaluates the performance of the Proximal Policy Optimization (PPO) algorithm in the General Video Game Artificial Intelligence (GVGAI) evaluation environment, which provides subdivision of a virtual game world into different phases or levels. Although PPO generally reports excellent results, it is notable that the algorithm suffers from overfitting to the training set.

As part of Pequi Mecânico — UFG's competitive robotics team — I collaborated on building autonomous systems for the Latin American Robotics Competition (LARC). Beyond the competitions, I helped bring robotics into the broader community through hands-on Arduino workshops.

• Designed navigation control pipelines for the IEEE SEK team, winning 1st place at LARC 2017
• Built navigation and control pipelines for the IEEE VSSS robot soccer team
• Organized workshops introducing Arduino and robotics fundamentals to the local community

Taught technology classes to high school students, introducing them to programming and robotics fundamentals. The curriculum included hands-on projects, such as building Android apps with MIT App Inventor, creating games with Stencyl, and preparing robots for the Brazilian Robotics Olympiad (OBR).