Վաղը կարևոր սեմինար է պլանավորման ուսուցման վերաբերյալ
Arec Jamgochian
Learning Safe Plans Under Uncertainty.
Friday, August 23, 2024 | 4:30 PM - 6:00 PM | YSU Krisp-AI Lab
Abstract:
Can LLMs plan? How reliable are robot policies that are learned end-to-end? Will agentic systems generalize given internet-scale data? All important questions… but not really one’s I’ll answer (maybe at the end). This talk will focus instead on a more bottoms-up perspective of data-driven planning — how to use good predictive models of a structured system’s behavior to come up with a sequence of actions that will optimize some objective. I’ll talk about types of uncertainty present in agentic systems, frameworks for handling uncertainty, and some roles for learning when planning (think reinforcement learning, imitation learning, AlphaZero, etc.). I’ll then overview some of our recent work on ConstrainedZero (IJCAI 2024), where we learn to plan under imperfect information while imposing safety constraints. Finally, I’ll briefly discuss how we are using some of these methods at TerraAI to do sustainable mineral and geothermal exploration.
About Arec:
Arec recently finished his PhD in autonomy (Aero/Astro + CS) from Stanford University with Prof. Mykel Kochenderfer. His research investigated methods for safe, data-driven, autonomous planning under uncertainty, with special applications to self-driving cars and energy systems. He just began a role as a founding research scientist at TerraAI, a Khosla-backed start-up working on AI for sustainable energy and resource planning. He has worked in autonomous driving, rocketry, and in finance at BlackRock AI Labs with Prof. Stephen Boyd. He was the head teaching assistant for Stanford CS classes on optimization and sequential decision-making, as well as an NSF graduate research fellow and an Accel fellow.
