The project supports acquisition of a hardware cluster and development of software frameworks to support Monte Carlo methods in articial intelligence, for tasks such as model compilation through machine learning over the results of Monte Carlo simulation, application of Monte Carlo methods to solve single-agent and multi-agent sequential decision making problems, and integration of machine learning methods into Monte Carlo search to improve real-time decision making. These software frameworks will include implementations of baseline and state-of-the-art algorithms for each task, with a goal of release with generic APIs and open-source availability to make it easy for other researchers to add new methods and to connect external simulators to the frameworks.
The algorithms and tools have many important applications, including (a) optimization of forest management to jointly minimize the risk of catastrophic fires and maximize biological and economic benefits, (b)design and validation of multiagent control methods for reducing congestion in air traffic control, (c) design and validation of multiagent control methods for micro air vehicles, and (d)modeling spatio-temporal distribution of species to support management of endangered and threatened species. The hardware cluster and software frameworks will be integrated into the undergraduate and graduate curriculum at Oregon State University, as well as various outreach beyond Oregon State.
Monte Carlo methods for artificial intelligence carry out complex tasks by applying randomized search techniques. This project developed Monte Carlo algorithms for planning, scheduling, and probabilistic inference. These algorithms were implemented in computer programs and incorporated into a software package for Monte Carlo Planning, which may be downloaded from http://web.engr.oregonstate.edu/mcai/materials/. To develop these methods and apply them to large-scale artificial intelligence problems, this project purchased a 28-node high-performance compute cluster. Each node has two 6-core processors, for a total of 336 cores. Six additional nodes were purchased with funds provided by other researchers at Oregon State University. The Monte Carlo methods and the compute cluster have been applied to several challenging problems including (a) fitting flexible models of the spatial distribution of birds, moths, and plants at various locations in North America and Australia, (b) planning the purchase of conservation reserves to protect the Red Cockaded Woodpecker in the southeastern US, (c) planning the allocation of conservation reserves for plants in the state of Victoria, Australia, (d) planning the siting and coordination of emergency vehicles for the Corvallis Fire Department, (e) computing optimal management strategies for managing invasive plant species in river networks, (f) computing optimal management strategies for deciding when to suppress lightning-caused wildfires in Eastern Oregon ponderosa pine forests, (g) reasoning about multiple phrases in natural language that refer to the same entity, (h) detecting complex, articulated objects against noisy backgrounds in computer vision, (i) detecting novel objects in computer vision, (j) detecting insider threats in organizations, (k) planning multi-agent coordination methods for the US air traffic control system, and (l) modeling the behavior of migrating birds over the eastern US and Canada. This project also funded two one-week courses in Monte Carlo Artificial Intelligence, which were held in Corvallis Oregon in March of 2012 and 2013. A total of 41 undergraduate students participated in these classes. All course materials from the class are available at http://web.engr.oregonstate.edu/mcai/. The course materials introduce Monte Carlo methods by showing how they can be applied to games and puzzles such as Connect-4, Yahtzee, Backgammon, Havannah, and Clue as well as to a bird conservation planning problem. This project contributed new theoretical insights into the factors that affect the success of Monte Carlo tree search algorithms and into the soundness of various forms of abstraction for accelerating Monte Carlo tree search. New algorithms for Monte Carlo solution of Markov Decision Problems were discovered that require significantly fewer Monte Carlo samples than previous methods. This project contributed to the publication 27 refereed articles and 2 doctoral dissertations.
