Evolving under tasks of incomplete information: streaming and self play

只有部分可观察到的任务对机器学习，特别是遗传编程(GP)提出了独特的挑战。部分可观察到的应用的概念将从与伙伴组织相关的两个具体应用背景来考虑。第一种情况是在有限的标签可用性下的流数据分析。目标是从代表任务当前状态的标记数据的某个小子集开始，然后能够识别生成数据流的非静态过程的基本性质何时发生了变化。在识别到流的行为中的这种变化时，当前的GP解决方案应该选择性地请求对应于该变化的数据的标签。然后需要做出决定，以确定应更新或替换哪些当前受雇于检测流变化/标签的全科医生。 第二个应用情境是学习在不完全信息博弈下通过自我游戏来构建非玩家角色(NPC)。NPC是基于计算机的实体，它出现在游戏中，为人类玩家提供额外的兴趣。在这项工作中，我们特别感兴趣的是出现在扑克游戏中的NPC的案例。然而，我们的目标不是培养出尽可能强大的球员，而是提供与目前参赛的人类球员互补的球员。扑克代表了一个有趣的任务领域，因为它也是基于不完全信息(不知道其他玩家的牌)、随机(无法预测牌出现的顺序)和对手玩家的能力未知的。从机器学习的角度来看，扑克任务的随机性、缺乏完全信息以及其他玩家采用的策略的非平稳性都使得扑克任务特别具有挑战性。

Tasks that only are only partially observable present unique challenges for machine learning in general and genetic programming (GP) in particular. The notion of a partially observable application will be considered from two specific application contexts of relevance to the partner organization. The first context is that of streaming data analysis under limited label availability. The goal is to begin with some small subset of labelled data representing the current status of a task, and then be able to recognize when the underlying nature of the non-stationary process generating the data stream has changed. On recognizing such a change in the stream's behaviour, the current GP solutions should selectively request labels for the data corresponding to the change. Decisions then need to be made to determine which GP individuals currently employed for detecting change / labelling the stream, should be updated or replaced. The second application context is that of learning to construct non-player character (NPC) through self play under a game of incomplete information. A NPC is computer based entity who appears in a game to provide additional interest for human players. In this work we are particularly interested in the case of NPC who appear in the game of poker. The goal however, is not to produce the strongest possible player, but to provide players who are complementary to the human players currently participating. Poker represents an interesting task domain because it is also based on incomplete information (no knowledge of the other player's cards), stochastic (cannot predict the order of card apparence) and the capability of opponent players is unknown. The stochastic nature of the task, lack of complete information and non-stationary nature of strategies adopted by other players all make the poker task particularly challenging from the perspective of machine learning.