My current research interest is in multitask learning methods that could be implemented for long-horizon tasks in RL. I am currently working on improving a method that uses a Mixture of Experts that allows one high-level policy to maintain high performance with several tasks. Some research questions I'd like to explore with this topic include:

• Does MTRL improve performance in long-horizon tasks via dividing the tasks into sub-tasks? Has research shown this method to be promising so far?
• How important is using skill-discovery for defining sub-tasks in long-horizon tasks?

In the future, I may potentially also branch into skill-discovery or curriculum learning while exploring methods for agents to perform well in long-horizon tasks.