Training a Deep Learning-based Agent/Robot at Scale

National Tsing Hua University Prof. Min Sun

計畫主持人:國立清華大學 電機工程學系孫民教授

放眼下一波許多擁有重大商業價值的人工智慧系統,需要處理的工作(tasks)變異度將越來越大。依賴蒐集大量人工標注的訓練資料,再利用監督式學習 (Supervised Learning)的方式已經不敷使用。本計畫專注關鍵技術讓深度學習可規模化訓練複雜的智慧代理人/機器人,讓相關應用搶先於國際商業化發展。主要技術:虛擬環境中訓練、自主學習、增強學習。

In the next wave, numerous artificial intelligence systems with significant commercial value are requested to handle the growing variance of tasks. The method of relying on collecting a large number of manually labeled data before supervised learning is no longer sufficient. This project focuses on key technology to enable a deep learning-based agent/robot to train at scale, which can boost international commercialization of related applications. Features: training in virtual environments, self-learning, reinforcement learning.