Skill Acquisition Learning (SAiL)
This project aims to learn skills by mimicking experts’ behaviors in given tasks. The proposed SAiL engine is trained to perform action prediction tasks from demonstrations by learning a mapping function between observed states and actions. The main challenges in real applications, medical and health care for example, are that the collection of such experts’ demonstrations is very expensive and It takes a large amount of time and money for expert training. Additionally, the demonstration varies among different experts, which makes it hard to reuse and interpret. For each individual expert, his decision-making process can also be affected by different factors and may not be consistent or easy to explain.
SAiL provide the solutions to aforementioned issues from the following critical dimensions.
- Sample efficient and requires fewer human efforts to generate demonstrations. This is achieved by exploiting information from perfect and imperfect demonstrations.
- Offers interpretable results and provides reusable skills. This is achieved by providing user friendly output decisions which are easy to understand including important feature rankings for each decision and reusable skills. For example, the reusable skills could be different treatment plans for patients with different symptoms.