软件新技术与产业化协同创新中心

摘 要：

Deep neural networks have proven to be successful in many identification tasks, however, from model-based control perspective, these networks are difficult to work with because they are typically nonlinear and nonconvex. Therefore, many systems are still identified and controlled based on simple linear models despite their poor representation capability. In the first part of the talk, I will introduce our recent work on bridging the gap between model accuracy and control tractability faced by neural networks, by explicitly constructing networks that are convex with respect to their inputs. We show that these input convex networks (ICNN) can be trained to obtain accurate models of complex physical systems. Experiment results demonstrate the good potential of the proposed ICNN approach in a variety of control applications. In particular we show that in the MuJoCo robotics locomotion tasks, we could achieve over 10% higher performance using 5× less time compared with state-of-the-art model-based reinforcement learning method; and in the building HVAC control example, our method achieved up to 20% energy reduction compared with classic linear models. Besides the complexity of system dynamics, physical systems often encompass a myriad of uncertainties that come from human behavior and the environment. In the second part of the talk, I will introduce a work in collaboration with DeepMind on learning reinforcement learning policies that are robust to perturbations in the environment dynamics. We propose a new data-driven algorithm for incorporating robustness into standard RL called Data-Driven Robust Maximum a-posteriori Policy Optimization (DDR-MPO). This algorithm first learns several transition models from the environment perturbation datasets, then incorporating the transition models into the simulator as an uncertainty set. We show that DDR-MPO outperforms standard RL algorithms in a variety of MuJoCo domains under different perturbed environments.

报告人简介:

石媛媛 (Yuanyuan Shi) 于2011-2015年就读于南京大学工程管理学院自动化专业，获得工程学士学位。2015-2020年于美国华盛顿大学西雅图分校（University of Washington, Seattle）电子及计算机工程学院获得博士学位，导师为Baosen Zhang。目前在加州理工学院跟随Anima Anandkumar与Adam Wierman进行博士后研究。将于2021年入职加州大学圣地亚哥分校 (UCSD)电子及计算机工程系担任助理教授。她的主要研究方向为智能控制, 机器学习及其应用在智能电网，物流以及机器人系统。她已在IEEE Transactions on Automatic Control, IEEE Transactions on Power Systems等国际期刊以及ICLR, NeurIPS, CDC, ACC等机器学习和控制会议上发表多篇论文。曾在Google DeepMind, 京东北美研究院等多处实习，以及获得由MIT授予的2018女性学术新星奖 (Rising Stars in EECS)。

腾讯会议ID: 832 554 424