Title: The Last Four Frames Are Not All You Need: Learning to Behave in a Partially Observable World
Speaker: Leslie Pack Kaelbling
Abstract: All robotics problems are partially observable. In some problems, the partial observability is relatively superficial, enabling solution via planning and/or learning methods that assume full observability. The modern embodied intelligence research community spends most of its time looking under this fully-observable lamppost [1]. But many problems of significant importance are significantly partially observable.
We know that obtaining optimal (or even reasonably good) solutions to general partially observable Markov decision problems (POMDPs) is intractable or even undecidable. Is that a reason to ignore or give up on them?
I'll argue that it's not. The fact that humans and other animals can learn and behave effectively in extremely partially observable domains argues that there is at least a subclass of general POMDPs that can be solved efficiently. I am interested in finding the structures and regularities in the real physical world that render many partially observable problems tractable. This talk will be a combination of tutorial and speculation, hoping to incite discussion, and no actual recent research results.
[1] en.wikipedia.o...
Bio: Leslie Pack Kaelbling is the Panasonic Professor of Computer Science and Engineering at Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology, where she co-leads the Learning and Intelligent Systems (LIS) group. She is widely known for her research on Partially Observable Markov Decision Processes (POMDPs) in the context of artificial intelligence and robotics, as well as her work on reinforcement learning, belief-space planning, state estimation and integrated task and motion planning. Leslie has received the NSF Presidential Faculty Fellowship, the IJCAI Computers and Thought Award, as well as the Bose Award for Excellence in Teaching. She is also the founding editor-in-chief of the Journal of Machine Learning Research (JMLR), and a fellow of the Association for the Advancement of Artificial Intelligence (AAAI).
21 авг 2024