Introducing self-organizing neuro-fuzzy Q-networks
Trained with unsupervised learning and finished with an offline model-free fuzzy reinforcement learning approach called fuzzy conservative Q-learning (FCQL), this approach offers more effective and interpretable policies than deep neural networks, while facilitating human-in-the-loop design and explainability.
Code demonstrating our proposed method is publicly available.
☕️John Wesley Hostetter is the owner and developer of PySoft - a flexible Python Soft Computing library, as well as a Ph.D. student of Computer Science at North Carolina State University since the Fall of 2019. He received his Bachelor’s and Master’s degrees in Computer Science at Penn State and NC State University, respectively.
His research interests are concerned with how a mathematical field studying impreciseness called fuzzy logic may aid in re-imagining function approximation in artificial intelligence. More specifically, he develops alternatives to artificial neural networks that are more interpretable called self-organizing neuro-fuzzy networks. These self-organizing neuro-fuzzy networks have the same power as artificial neural networks, but can explain their decision-making. With his adviser, Dr. Min Chi, they leverage this novel and interpretable function approximation technique to improve undergraduate STEM education.
Mr. Hostetter has worked as a graduate teaching assistant in courses such as: Java Programming Concepts; Automata, Grammars & Computability; Automated Learning & Data Analysis. He now works as a teaching assistant for two courses at NC State’s Artificial Intelligence Academy: Introduction to Artificial Intelligence & Machine Learning, in addition to his graduate research assistant role.
Skills
Fuzzy sets, fuzzy logic, rough sets, approximate reasoning, impreciseness, vagueness, uncertainty
Supervised/unsupervised/reinforcement learning
Experience with PyTorch, Optuna, iGraph, etc.
Experience
Scholarships, Fellowships, Grants & Awards
Gallery
