CyberSEES: Type 2: Computing and Visualizing Optimal Policies for Ecosystem Management

许多生态系统表现出空间扩散的过程，我们希望管理，以促进或防止。例如，我们力求防止入侵物种的扩散，同时促进濒危物种的扩散。对于荒地火灾，我们寻求促进低强度地面火灾的蔓延，防止危险燃料的积累，同时防止破坏濒危物种栖息地和宝贵木材的高强度树冠火灾的蔓延。这些管理问题在数学上可以表述为在空间区域上定义的马尔可夫决策问题（MDP）。然而，由于这些问题的空间性质，MDP是巨大的，不能解决任何现有的算法。本计画针对空间MDP开发新的MDP求解演算法。这些算法将与生态系统模拟器一起工作（而不需要明确的模型），它们还将解决灾难性后果的风险，如物种灭绝或灾难性野火。为了弥合MDP的计算解决方案与政策制定者实际采用此类解决方案之间的差距，本研究开发了可视化和交互方法，这些方法将允许利益相关者（例如，政策制定者、土地所有者、木材行业代表、保护生物学家），以理解和批评问题的提法和由此产生的解决方案。更广泛的影响新的方法将在五个管理问题上进行测试：（a）在河流网络中传播的Tambirk，（B）在美国西部范围内传播的Cheatgrass，（c）在加州和俄勒冈州传播的橡树猝死，（d）决定何时让野火燃烧与抑制野火，以及（e）决定在景观中放置燃料减少处理以减少火灾风险。在Tamarak，Cheatgrass，和橡树猝死问题将在程式化的设置，相关的环境属性和成本可以变化的研究。这些研究的目的是了解不同的空间扩散过程（表现为这些不同的入侵物种）如何决定最佳管理政策的结构。研究结果将在自然资源经济学文献中发表，并与这些领域的决策者进行讨论。野火问题（“让燃烧”和“燃料处理”）将在一个真实的实验室进行研究--一个位于德舒特国家森林公园的公有场地，里面混合了黄松和黑松。一个与美国林务局合作的消防经理将招募一个利益相关者小组，使用我们将开发的可视化和交互工具来分析和评论拟议的管理政策。在这个项目中开发的问题和技术将成为第一个计算可持续性暑期学校的核心，这将由研究团队组织。研究结果也将被整合到OSU生态信息学暑期研究所（NSF REU网站）和OSU春假课程蒙特卡洛人工智能为大三学生。四个博士学生和一个博士后将在这个项目中进行培训。

Intellectual MeritMany ecosystems exhibit spatially spreading processes that we would like to manage to either promote or prevent. For example, we seek to prevent the spread of invasive species while promoting the spread of endangered species. For wildland fire, we seek to promote the spread of low-intensity ground fires that prevent the buildup of dangerous fuels while preventing the spread of high-intensity crown fires that destroy endangered species habitat and valuable timber. These management problems can be formulated mathematically as Markov Decision Problems (MDPs) defined over spatial regions. However, because of the spatial nature of these problems, the MDPs are immense and cannot be solved by any existing algorithms. This project develops new MDP solution algorithms for spatial MDPs. These algorithms will work with ecosystem simulators (rather than requiring explicit models) and they will also address the risk of catastrophic outcomes such as species extinction or catastrophic wildfires. To bridge the gap between the computational solution of an MDP and the actual adoption of such solutions by policymakers, this research develops visualization and interaction methods that will allow stakeholders (e.g., policymakers, land owners, timber industry representatives, conservation biologists) to understand and critique both the problem formulations and the resulting solutions. Broader ImpactsThe new methods will be tested on five management problems: (a) Tamarisk spread in river networks, (b) Cheatgrass spread in Western US range lands, (c) Sudden Oak Death spread in California and Oregon, (d) deciding when to let a wildfire burn versus suppressing it, and (e) deciding where to place fuel reduction treatments in the landscape to reduce fire risk. The Tamarisk, Cheatgrass, and Sudden Oak Death problems will be studied in stylized settings where the relevant environmental properties and costs can be varied. The goal of these studies is to understand how the different spatial spreading processes (exhibited by these different invasive species) determine the structure of the optimal management policy. The results will be published in the literature on natural resource economics and discussed with policymakers in these areas. The wildfire problems ("let burn," and "fuel treatment") will be studied in a real landscape-a publicly-owned site in the Deschutes National Forest containing a mix of Ponderosa and Lodgepole pine. A collaborating fire manager with the US Forest Service will recruit a panel of stakeholders to analyze and critique the proposed management policies using the visualization and interaction tools that we will develop.Problems and techniques developed in this project will form the core of the first Summer School in Computational Sustainability, which will be organized by the research team. The results will also be integrated into the OSU Summer Institute in Eco-Informatics (an NSF REU Site) and the OSU Spring Break Course in Monte Carlo AI for junior undergraduates. Four Ph.D. students and one Postdoc will be trained during this project.