Learning Complex Probabilistic Models from Data

点击翻译按钮获取中文摘要

Many problems in the real world involve systems that are complex, stochastic, and partially observable -- for example, jet engines, the human body, and the stock market. Recent developments in probabilistic representation and reasoning technology, especially probabilistic (Bayesian) networks, have led to a rapid growth of successful applications in engineering, finance, medicine, software systems, etc. This project concerns the development of techniques for automated construction of complex probabilistic networks directly from data. Previous work by the principal investigator has shown that distributed, gradient-based methods can be used to learn probabilistic networks with hidden variables. This work is being extended to handle models of temporal processes, models with mixtures of continuous and discrete variables, models with unknown structure, and models for which various types of prior knowledge are available, thereby providing more powerful learning methods for applications such as data mining, forecasting, speech recognition, and computer vision. Because probabilistic networks can provide compact and easily understandable models of very complex processes, the project will also yield new tools for automated scientific investigation in cases where current data analysis tools are ineffective -- for example, in understanding the complex pathways by which DNA lesions lead to tumors.