| Title | : | Transfer Learning and Explainability for Safe Offline RL |
| Speaker | : | Richa Verma (IITM) |
| Details | : | Wed, 14 Jun, 2023 11:00 AM @ Aryabhatta Hall, CSB |
| Abstract: | : | Offline RL is a practical paradigm where an agent is trained solely using a fixed dataset of trajectories without engaging in interactions with the environment during the learning process. The ability to learn from such a fixed dataset is essential for developing scalable and adaptable data-driven learning techniques. It is also of paramount importance in scenarios where accessing the environment during training is costly or unfeasible. Safety is a key aspect to ensure that the policies learned using such techniques are deployable. We aim to study safety in offline RL through the lens of transfer learning and explainability. First, we focus on enhancing safety when there is an increased risk of the agent selecting potentially unsafe actions, particularly in the regions of the state space which have limited representation within the training dataset. Next, we plan to develop an explainability framework for safe offline RL algorithms, enabling end-users to gain more insight into the decision-making process of the agents trained using such methods. Further, following th e recent increase in literature using self-supervised learning for offline learning, we also discuss the possibility of generating post-hoc explanations for policies trained using such methods. |
