Mark Crowley, Oregon State University
Big Data and Big Decisions - From Sustainable Management to Large Scale Spatiotemporal Decision Making
Advances in recent years in computing power, storage and communication have led to an explosion of accessible data about our world. Meanwhile, advances in machine learning and data analytics have just begun to allow us to begin sensibly dealing with all this data. However, the end goal of most analytics applications is not the analysis itself, but rather to provide an aid to human decision makers in large scale sequential decision making problems. The goal of my research is to improve our ability to provide tools which automate some of this decision making. Sustainable management problems such as Forest Management and Invasive Species Control provide good examples of the challenges that can arise in such Big Data-Big Decision problems. One challenge is that acquiring new data to evaluate a policy, usually through large black box simulations, can be very slow, hindering simulation based optimization schemes. Another challenge is the presence of spatial or relational structure within the decision problem which leads to a combinatorial explosion of complexity.
In this talk I will present two algorithms which make significant progress on these challenges based on work I've done with collaborators in Forestry and Resource Economics. The first algorithm, DDV, provides confidence interval guarantees on convergence to the optimal policy using a much more efficient sampling approach than previous model-based planning methods. The second algorithm, Equilibrium Policy Gradients, is a direct policy search algorithm that is the first approximate MDP solver that can handle spatial structure in large action spaces. I will conclude with a brief discussion of some of my ongoing work and future research plans pursuing new questions raised by these projects, as well as how my algorithms generalize to other Big Data-Big Decision domains such as advertising in social networks, infectious disease control and network optimization.
Mark Crowley is a Postdoctoral Scholar in the Machine Learning group in the Department of Electrical Engineering and Computer Science at Oregon State University. He received his Ph.D. and M.Sc. in Computer Science from the University of British Columbia working in the Laboratory for Computational Intelligence, and a B.A. in Computer Science from York University in Toronto. Before graduate school he worked for six years in software development, four at the IBM Toronto Lab. His research focuses on algorithms, tools and theory at the intersection of Machine Learning, Optimization and Probabilistic Modelling. In particular he has helped make significant advances in the problem of Large Scale Spatiotemporal Decision Making. He often works in collaboration with researchers in other fields such as sustainable forest management, ecology and resource economics. He is an active part of building the interdisciplinary Computational Sustainability community which connects research in the computational sciences with sustainable management and analysis problems.