CAREER: Harnessing the data revolution for predicting and managing ecosystem regime shifts

生态系统的突然变化不仅代表了我们环境中一些最复杂和最有影响的变化，也是最难以预测和管理的变化。甲虫和火灾对森林的破坏，大西洋鳕鱼渔业的崩溃，或者疾病的爆发，都可能是这种突然变化的例子。生产力和危机之间的突然转变损害了对海洋和森林的有效管理。从卫星和微型传感器到大型天文台，我们收集数据的方式发生了一场革命，这将使新的渴望数据的机器学习方法在这些范围内进行预测和管理是可行的，但清晰度的净收益仍是未知的。这项研究旨在评估机器学习和人工智能工具如何提高预测和管理生态系统突然变化的能力，并了解无法做到的限制。使下一代生态学家和环境科学家能够充分了解这些工具，从而做出明智的决定，这是实现这一愿景的关键。一个综合的研究和教育计划将通过创新的教学方法来解决这些问题，该方法寻求促进数据科学与生态和环境问题之间的多样性。这项研究寻求通过调整和结合机器学习算法与机械驱动的模型和新兴的生态数据源来促进生态预测和决策方面的现有知识。该项目的第一阶段评估了递归神经网络结构预测生态系统动态的有效性，这些生态系统能够发生突然的制度变化--理论表明，现有的机器学习方法可能会失败。然后，这项研究通过将机器学习方法与生态学理论指导下的机械模型相结合，寻求更稳健的预测设计。该项目的第二阶段寻求借鉴强化学习中的新方法，以解决养护方面的常见优化问题，例如确定可持续的收获水平或保护区的位置。在这里，研究将把“深度”强化学习的最新发展与基于过程的生态管理方法结合起来。这项研究的两个阶段都将得到开发开放源码软件工具的支持，以便在各种情况下实施这些方法。该项目的结果，包括所产生的科学出版物和软件产品的更新和链接，可以在https://carlboettiger.info.This上找到，该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Abrupt ecosystem shifts represent not only some of the most complex and impactful changes in our environment, but also the most difficult to predict and manage. Forest devastation by beetles and fire, the collapse of the Atlantic cod fishery, or the outbreak of a disease may all be examples of such sudden changes. Effective management of oceans and forests is impaired by abrupt shifts between productivity and crisis. A revolution in how we collect data, from satellites and micro-sensors to large-scale observatories will make new data-hungry machine learning approaches to forecasting and management across these scales feasible yet the net gain in clarity remains unknown. This research seeks to evaluate how the tools of machine learning and artificial intelligence can improve the ability to predict and manage sudden ecosystem change, and understand the limits where it cannot. Empowering the next generation of ecologists and environmental scientists to understand these tools sufficiently to make informed decisions is key to realizing this vision. An integrated research and education program will tackle these questions through an innovative pedagogical approach that seeks to promote diversity at this interface between data science and ecological and environmental issues.This research seeks to advance current knowledge in ecological forecasting and decision-making by adapting and combining machine-learning algorithms with mechanistically motivated models and emerging ecological data sources. The first phase of the project assesses the effectiveness of recurrent neural network architectures to predict dynamics in ecological systems that are capable of sudden regime shifts – a setting where theory suggests existing machine learning approaches are likely to fail. This research then seeks more robust forecast design by combining machine learning approaches with mechanistic models guided by ecological theory. The second phase of the project seeks to draw on emerging methods in reinforcement learning to address common optimization problems in conservation, such as determining sustainable harvest levels or the location of protected areas. Here, research will blend recent developments in “deep” reinforcement learning with process-based approaches to ecological management. Both phases of this research will be supported by the development of open source software tools for implementing these approaches in a wide variety of contexts. Results of the project, including updates and links to resulting scientific publications and software products can be found at https://carlboettiger.info.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.