Bayesian Transfer Learning
Abstract
Bayesian Transfer LearningHowie Choset February 8, 2020AbstractBoth human and non-human animals when faced with a new, albeit similar, task, do not have to Òre-learnÓ from scratch how to perform a task. Standard machine learning techniques tend to not support this capability because once a policy is trained, it is a solution for one isolated task. We believe, however, that many tasks share commonalities and therefore provide information can be leveraged when faced with similar, yet new, tasks. This is called transfer learning. Our proposed method for enabling transfer of knowledge among tasks is based on hierarchical Bayesian learning. Typical Bayesian learning maintains a belief about the parameters that govern the data distribution for a particular task. In our proposed approach, rather than learning a set of parameters that govern a belief for a single task, we consider a group of related tasks that have been sampled from the same underlying task distribution. We are then able to capture dependencies among tasks by imposing a common prior over task-specific parameters. Unlike typical approaches that use a generic, uninformed prior, the prior in our case is also learned over the task distribution, allowing this prior to represent the patterns that are common among these related tasks. This informed prior allows an agent to begin the learning process on a new task with a belief about the task based on past experiences, which it can then quickly refine as the learning process continues. In the proposed work, we will show that many tasks of interest, to both the robotics and machine learning communities, have enough in common that large amounts of knowledge can be transferred among them to significantly accelerate the learning process. Equivalently, we also show that experiences of an agent learning typical robotics tasks will contain a significant amount of knowledge that is applicable to future tasks. Ultimately, we seek to automatically learn relationships among related tasks in an information-theoretically principled way to generalize among a broad, yet related, set of tasks.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 20, 2022
- Source ID
- W911NF2210252
Entities
People
- Howard Choset
Organizations
- Army Contracting Command
- Carnegie Mellon University
- United States Army