Towards Building Autonomous Agents that Handle Open-World Environments
Abstract
Quals talk:
Artificial Intelligence agents have shown exceptional performance
in
many closed worlds
domains such as
games where the action space, state space, and the transition dynamics
are fixed for duration of the task. Even
minor changes in the environment dynamics, however, can lead to
catastrophic results for closed-world agents. To
make the agents more suitable for a real-world setting, we need to
relax
closed-world assumptions and make agents
robust to novel, unseen, and unexpected scenarios. To build agents
more
robust to unseen scenarios we propose (1) A
novel experimental environment called
NovelGridworlds
to
benchmark
and develop such agents which can respond
to novelties. (2) A novel architecture called RAPid-Learn which
synergistically integrates planning and learning
methodologies to learn to best respond to the sudden and unexpected
changes in an agent's environment (i.e.,
novelties). In this talk, I will present the details of the proposed
experimental environment and show results of
the proposed architecture. I will also discuss the improvements and
future work of this research.
Research area: Reinforcement Learning, Robotics, Symbolic Learning
Please join meeting in JCC 435 or join via Zoom.
Join Zoom Meeting: https://tufts.zoom.us/j/8729644907
Password: See colloquium email
Dial-in not an option for this event.