Towards Fully Integrated Deep Learning and Reinforcement Learning for General Spatial Domains.

点击翻译按钮获取中文摘要

Artificial intelligence (AI) and machine learning (ML) research develops techniques that enable automated prediction, data analysis and decision making using diverse data sources. The areas of application for AI/ML technologies are almost limitless and have the capacity to radically impact human society and almost every industry in the coming years. Before an AI/ML system can be used, it must be “trained” using relevant data from which it “learns” patterns etc. Two subfields of study that examine how to best train AI/ML systems are Deep Learning (DL) and Reinforcement Learning (RL). Thus far, DL and RL have been applied primarily to the analysis of photos, videos and video games but this omits patterns that occur in many important real world scenarios. For example, analyzing the spread of a forest fire involves tracking the temporary condition of burning across a landscape as it responds to vegetation, moisture, and altitude but irrevocably alters other vegetation, soil, and other local features. A Spatially Spreading Process (SSP) is said to exist in a domain when some local features change over time and space in this way based on proximity among objects but without resulting from the movement of an object or objects. It is a local change of some properties across a spatial landscape. Recently a convergence DL and RL research has begun to offer the potential for a quantum leap in the ability to use AI/ML for predicting/analyzing SSP data. However, a number of gaps remain which the proposed research program will investigate via three objectives: (1) develop a novel Deep RL framework that integrates actions directly into the basic calculations of deep neural networks for the purposes of prediction and decision making in 2D spatial environments, (2) investigate predictive classification for non-visual, spatial 3D environments using existing methods and a novel policy approach, and (3) formulate a more general theory of data augmentation for DL and RL algorithms that accounts for SSPs in both 2D and 3D environments.The results of this five year program will provide a base for the longer term goal of augmenting human decision making in a transparent and dependable way for problems with complex spatiotemporal dynamics and large-scale streaming data. The accuracy and efficiency of our methods will be validated on a range of domains with complex spatiotemporal structures including forest wildfire management, flood prediction and classification of Alzheimer's disease based on 3D diffusion MRI data. This program will contribute significantly to the growing footprint of forward-thinking AI/ML research coming out of universities in Ontario and Canada. The program will train ten graduate and five undergraduate students in rigorous research skills and the use of AI/ML software tools increasingly in demand within the information economy.