Evaluating Next Generation Probabilistic Planners

本研究计划将发展规划演算法及一套评估概率规划者的一般方法。概率规划是顺序决策制定的领域，涉及在可用的操作符具有不确定的结果时选择改变世界状态的操作符。这个项目的驱动目标是推进概率计划的艺术状态，以提高效率，改进对问题变化的鲁棒性，并扩大对现实世界问题的适用性。为了实现其目标，该项目将侧重于两个相互关联的任务。首先，它将提出并发展一种评估概率规划者的方法。这将需要研究一组替代方案，并运行实验，将评估指标与日益现实的领域中的理想结果联系起来。该项目的努力将通过两年一次的国际规划竞赛（IPC）与更大的研究界密切协调，该竞赛将很快在其现有结构中引入概率跟踪。该项目将组织跟踪，并将为社区提供一套用于执行和评估概率领域计划的软件程序。其次，项目成员将继续开发他们自己的规划算法，特别强调利用概率规划和强化学习之间关系的方法。该项目将在研究概率规划问题和如何衡量该领域的进展方面产生明确的研究影响。它还将通过探索基于实例的学习技术来更有效地学习计划，从而推进规划和强化学习方面的最新技术。然而，大部分工作的重点将放在其对整个规划社区的广泛影响上，包括具体的领域描述语言、评估软件和基准问题，这些问题将有助于将社区的努力集中在开发算法上，以解决重要的科学和经济利益问题。

This research project will develop planning algorithms and a set of general methods for evaluating probabilistic planners. Probabilistic planning is the area of sequential decision making concerned with choosing operators that change the state of the world when the available operators have uncertain outcomes. The driving goal of this project is to advance the state of the art of probabilistic planners toward increased efficiency, improved robustness to problem variations, and broadened applicability to real-world problems. To accomplish its goal, the project will focus on two interrelated tasks. First, it will propose and develop a methodology for evaluating probabilistic planners. This will require studying a set of alternatives and running experiments to correlate evaluation metrics with desirable outcomes in increasingly realistic domains. The project efforts will be coordinated closely with the larger research community through the biannual International Planning Competition (IPC), which will soon introduce a probabilistic track to its existing structure. This project will organize the track and will provide the community with a set of software programs for executing and evaluating plans in probabilistic domains. Second, the project members will pursue the development of their own planning algorithms, with a particular emphasis on approaches that exploit the relationship between probabilistic planning and reinforcement learning.The project will have definite research impacts in its study of the problem of probabilistic planning and how progress should be measured in the field. It will also advance the state of the art in planning and reinforcement learning through the exploration of instance-based techniques for learning to plan more effectively. However, the focus of the majority of the work will be on its broader impacts on the planning community as a whole, with concrete domain description languages, evaluation software, and benchmark problems that will serve to focus the community's efforts toward developing algorithms to solve problems of significant scientific and economic interest.