Assessing AI systems is difficult. Humans rely on AI systems in increasing ways, both visible and invisible, meaning a variety of stakeholders need a variety of assessment tools (e.g., a professional auditor, a developer, and an end user all have different needs). We posit that it is possible to provide...
It is common practice in the unsupervised anomaly detection literature to create experimental benchmarks by sampling from existing supervised learning datasets. We seek to improve this practice by identifying four dimensions important to real-world anomaly detection applications --- point difficulty, clusteredness of anomalies, relevance of features, and relative frequency of...
Society faces many complex management problems, particularly in the area of shared public resources such as ecosystems. Existing decision making processes are often guided by personal experience and political ideology rather than state-of-the-art scientific understanding. This dissertation envisions a future in which multiple stakeholders are provided with computational tools for...
This work is inspired by problems in natural resource management centered on the challenge of invasive species. Computing optimal management policies for maintaining ecosystem sustainable is challenging. Many ecosystem management problems can be formulated as MDP (Markov Decision Process) planning problems. In a simulator-defined MDP, the Markovian dynamics and rewards...
Monte Carlo tree search (MCTS) is a class of online planning algorithms for Markov decision processes (MDPs) and related models that has found success in challenging applications. In the online planning approach, the agent makes a decision in the current state by performing a limited forward search over possible futures...
Machine learning systems are generally trained offline using ground truth data that has been labeled by experts. However, these batch training methods are not a good fit for many applications, especially in the cases where complete ground truth data is not available for offline training. In addition, batch methods do...
Many important application problems in engineering can be formalized as nonlinear
optimization tasks. However, numerical methods for solving such problems
are brittle and do not scale well. For example, these methods depend critically
on choosing a good starting point from which to perform the optimization search.
In high-dimensional spaces, numerical...
Knowledge compilation improves search-intensive problem-solvers that are easily specified but inefficient. One promising approach improves efficiency by constructing a database of problem-instance/best-action pairs that replace problem-solving search with efficient lookup. The database is constructed by reverse enumeration- expanding the complete search space backwards, from the terminal problem instances. This approach...
Remote sensing is the most practical way to acquire large amounts of land cover data for monitoring and understanding environmental change, so it is important to be able to map land cover from imagery. Maps defining land cover patches as polygons rather than pixels greatly improve processing efficiency in models...
Recent work has shown that AdaBoost can be viewed as an algorithm that maximizes the margin on the training data via functional gradient descent. Under this interpretation, the weight computed by AdaBoost, for each hypothesis generated, can be viewed as a step size parameter in a gradient descent search. Friedman...