Exploration and Learning in Heuristic Search

计算机科学的研究朝着解决更复杂、更困难的现实问题的目标发展。智能自动化决策需要对应用领域进行建模，并处理未来可能选择的潜在巨大空间。与我的研究小组和同事一起，我研究解决困难决策问题的高效搜索算法。目前的提案主要关注两个主题：在大型搜索空间中进行有效探索的问题，以及机器学习方法的使用。这些主题已经成为推动我的团队在不同应用领域的近期工作的主要共同主题。在过去的五年里，我的研究领域发生了根本性的变化。将深度强化学习与蒙特卡罗树搜索相结合的系统在围棋、国际象棋和将棋等复杂游戏中取得了超人的表现。DeepMind的Alpha Zero系统已经学会了从零开始玩这类游戏，没有任何关于游戏策略的人工输入。这些系统的一个美妙之处在于它们结合了学习和搜索的方式。他们创造了一个良性循环，机器学习改善了搜索过程，搜索也改善了学习。尽管这些算法取得了令人印象深刻的成功，但许多实际和基本性质的问题目前限制了它们的更广泛使用。一个主要的实际问题是需要大量的资源来训练对所学知识进行编码的大而深的神经网络。更基本的问题包括：如何控制搜索过程？当我们没有一个完美和有效的问题模型时，如何推广这些方法？在今后与学生和同事的工作中，我希望深入研究以下几个主题：1。继续研究启发式搜索中的探索性2。将我们的方法扩展到不太明确的游戏以外的问题。研究由于其特殊的数学结构我们知道“真实结果”的情况下的学习和搜索。为了研究这些研究问题，我计划继续研究构成重大挑战的具体应用。我想继续构建完整的高性能系统，并在标准基准和比赛中对它们进行测试。对这些方法的深入理解可能会导致进一步显著改进决策系统，它可以更好更快地搜索和学习，并可用于定义不太明确的问题。

Research in Computing Science progresses towards the goal of solving ever more complex, difficult real-world problems. Intelligent automated decision-making requires modelling an application domain, and processing a potentially huge space of possible future alternatives. With my research group and my colleagues I study efficient search algorithms for solving hard decision-making problems. The current proposal focuses on two topics: the question of efficient exploration in large search spaces, and the use of machine learning methods. These topics have emerged as big common themes that drive much of my group's recent work in a diverse set of application areas. My research area has radically changed over the last five years. Systems that combine deep reinforcement learning with Monte Carlo Tree Search have achieved super-human performance in complex games such as Go, chess and shogi. DeepMind's Alpha Zero system has learned to play such games from scratch, without any human input regarding playing strategy. A beautiful aspect of these systems is the way in which they combine learning and search. They create a virtuous cycle where machine learning improves the search process, and the search also improves the learning. Despite the impressive successes of these algorithms, a number of problems of both practical and fundamental nature currently limits their more widespread use. A major practical problem is posed by the massive resources required to train the large and deep neural networks which encode the learned knowledge. More fundamental questions include: how to control the search process? And how to generalize such approaches when we don't have a perfect and efficient model of a problem? In future work with my students and colleagues, I want to study the following topics in depth: 1. Continue the study of exploration in heuristic search 2. Extend our methods to problems beyond games, which are less well specified 3. Study learning and search in cases where we know the "true result" due to their special mathematical structure To study these research questions, I plan to continue working on concrete applications which pose significant challenges. I want to continue building complete high performance systems, and test them on standard benchmarks as well as in competitions. A deeper understanding of these methods will likely lead to further significantly improved decision-making systems, which can search and learn better and faster, and can be used for less well-defined problems.