II-EN: Software Tools for Monte-Carlo Optimization

点击翻译按钮获取中文摘要

The Computing Research Infrastructure project supports the development of an open-source software library for Monte Carlo methods in artificial intelligence on a cloud-based platform. Monte Carlo methods are randomized numerical algorithms used in AI, machine learning, data mining, and the physical sciences. As data size and model complexity continue to grow, advanced large-scale Monte-Carlo techniques (including parallel implementation) has become ubiquitous. However software tools to easily implement advanced techniques for large-scale Monte Carlo are not established or broadly available.The software library developed in this project will help bridge this gap, and lower the barrier to adoption of advanced Monte Carlo techniques by a broad research community. The library will include a variety of existing state-of-the-art algorithms, as well as novel software components. The algorithms and tools have many important applications, including: (a) optimization of ecological management problems, including endangered species conservation, forest fire management, and invasive species management (b) automated software testing, (c) optimization for experimental design in science and engineering, (d) tracking of multiple objects from noisy visual evidence, and (e) activity and object recognition in computer vision. These problems have significant societal and economic importance and the research has the potential to significantly extend current capabilities.The algorithms and tools will be implemented using a common interface supporting a cloud-based platform, which will allow other researchers to extend and apply the library to important applications. The software library will be integrated into theundergraduate and graduate curriculum at Oregon State University. In addition, an online course centered around the theory and application of the library components will be developed, facilitating use by a wide audience.The developed library will include components based on the investigators' research that realize a number of technicalinnovations for Monte-Carlo Optimization (MCO), including: a) Exploiting multi-fidelity simulators in MCO for offline and online planning, (b) Developing MCO techniques for item discovery problems, (c) Developing MCO techniques for online policy improvement in sequential decision making, (d) Learning to reduce branching factors for more efficient online MCO, and (e) Integrating symbolic reasoning and MCO for scalable sequential decision making. These new capabilities will advance the state-of-the-art in artificial intelligence and enable new applications to be addressed that are beyond the scope of prior MCO methods.