Transfer learning for continual learning in non-stationary environments

点击翻译按钮获取中文摘要

Currently available artificial intelligence algorithms require re-training whenever their application environment undergoes fundamental changes and a new task is given. The ability to build an automated algorithm that can learn reliably under changing situations will therefore set new standards in the field. The main goal of this project is to develop new frameworks in which reinforcement learning algorithms can perform more stably and achieve higher sample and computational efficiency among a set of similar tasks in stochastic environment. That is, we intend to use transfer learning to address continual learning in non-stationary environments. To do so, we propose the following three objectives: 1) We will build adaptive neural network-based reinforcement learning models that will construct a knowledge representation that fits the non-stationary environmental complexity. This can be understood as a particular type of continual learning. 2) We will evaluate the performance of the developed frameworks/models in real-world applications which will be identified in collaboration with LG Electronics Canada. The identified problems may involve vision inspection data as inputs. 3) We will investigate how these frameworks/models perform when transferred to new environments in terms of dynamics, rewards and goals. The transfer learning frameworks/models will be combined with the developed continual learning frameworks/models. To sum up, the proposed project is to advance the current state-of-the-art of reinforcement learning so as to improve the sample and computational efficiency across similarly configured tasks and to robustify its performance in non-stationary environment.