Efficiently solving hard optimization problems is critical in many disciplines such as scheduling, hardware design, code compilation and so on. These problems are often NP-hard and rely on search techniques to find good solutions. To make the search efficient, traditional techniques often rely on human experience to design search heuristics. In this talk, I will cover our recent works that learn to focus on important regions in order to make search more efficient. This can be achieved by either training a reinforcement learning agent if samples are abundant, or in the case of limited samples, learning a simple model on the fly. The application includes neural architecture search, network optimization, vehicle safety design, single and multi-objective black-box optimization, and so on.
主题：Solving optimization problems with learning-guided search
嘉宾：田渊栋博士，脸书（Facebook）人工智能研究院研究员及经理
时间：2021 年 10 月 16 号（周六）北京时间 早10：00
录播：
https://www.bilibili.com/video/BV14g411L7rL
嘉宾介绍：
田渊栋博士，脸书（Facebook）人工智能研究院研究员及经理，研究方向为深度强化学习，表示学习和优化。曾获得2021年国际机器学习大会（ICML）杰出论文奖提名（Outstanding Paper Honorable Mentions)，及2013年国际计算机视觉大会（ICCV）马尔奖提名（Marr Prize Honorable Mentions）。2013-2014年在Google无人驾驶团队任软件工程师。2005年及08年于上海交通大学获本硕学位，2013年于美国卡耐基梅隆大学机器人研究所获博士学位，围棋开源项目ELF OpenGo项目中研究及工程负责人和第一作者。
分享内容：
以上是我们对本次直播展示 PPT 的部分截取。后台回复关键词“
ORtalk
”可获取本期 PPT 的完整内容。
