- Frazier Baker - bakerfn@mail.uc.edu
- Jeremiah Greer - greerji@mail.uc.edu
- Dr. Fred Annexstein - annexsfs@mail.uc.edu
Reinforcement learning often involves having a machine "learn" multiple tasks, generally requiring rewards for each task. This is formulated as multi-objective optimization.
We seek to find a novel multi-objective optimization method for use in reinforcement learning.
Current methods for multi-objective optimization involve linear combinations of the reward terms; however, balancing each of the rewards has proven difficult.
OpenAI.com has suggested that Filter Methods may be useful for multi-objective optimization and have been studied very little in relation to Reinforcement Learning.
In addition, there are many other search or optimization methods which could be derived from other subfields of Artificial Intelligence that have yet to be explored.
see Jeremiah's Bio
- Developed Data Analysis tool for World Bank in first co-op
- Currently taking Intelligent Data Analysis
- Currently taking Machine learning (Ng) & Neural Network (Hinton) courses on Coursera
- Attended Deep Learning Seminar with Dr. Annexstein
see Frazier's Bio
- Experience in Bioinformatics applications of Data Analysis and Machine Learning
- Multiple journal publications and conference proceedings.
- Experience in Computational Physics Research with Dr. Gabriela Popa at Ohio University Zanesville.
- Multiple conference proceedings.
- Took CS 6052: Intelligent Data Analysis from Dr. Raj Bhatnagar in Fall of 2015.
- Currently taking CS 4033: Artificial Intelligence
- Participated in Independent Study on Deep Learning with Dr. Fred Annexstein in Spring 2017.
- Currently taking Parallel Computing with Dr. Fred Annexstein, exploring CUDA and distributed computing.
Our plan is to explore the use of filter methods for multi-objective optimization and reinforcement learning.
In addition, we will explore other areas of AI, particularly Genetic Programming, to find new methods of feature selection.
Next, we hope to test both of these using the publicly available OpenAI Gym MuJoCo environments.
Finally, we hope to explore applications to other datasets that would demonstrate real-world benefits, as time permits.