Guided Reinforcement Learning for Model Repair

The complexity of model repair often necessitates automation. With the growing demand for Machine Learning based solutions, Reinforcement Learning (RL) has seen use for model repair. However, RL has a key limitation - training takes (too) long. This limitation is intrinsic to the RL process, thus to overcome it, I had to gain a deep understanding of RL to pinpoint a solution to improve performance radically.

My solution is to supplement RL with human intelligence in an approach called Opinion-Guided Reinforcement Learning. Opinions are mathematical constructs defined by the framework of Subjective Logic, which allow users to express their belief and uncertainty. Through experimentation, I found that human guidance in the form of opinions, even uncertain ones, improves RL performance.

Currently, I am working to apply this approach to model repair as seen by my submission to the ACM SRC at MODELS 24, Towards Model Repair by Human Opinion–Guided Reinforcement Learning.