Coopetition: Cooperation in Competition and Its Application to Artificial Intelligence

合作：竞争中的合作及其在人工智能中的应用

• 批准号: 2115277
• 金额: --
• 依托单位: University of Southampton
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2018
• 资助国家: 英国
• 起止时间: 2018 至 无数据
• 项目状态: 已结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2115277
• 关键词: Coopetition; Cooperation; Competition; Its; Application

Consider a decentralised machine learning problem in which two rival online marketing companies aim to share their data in order to train a recommendation algorithm to maximise its accuracy level, as none of them has sufficiently large amount of data to do so. The more data they decide to share, the better improvement in the performance of the algorithm they can get. However, as they are also rivals, sharing too many data with each other will also induce a cost (e.g., losing sensitive data and market to the rival). On the other hand, without cooperation, they would not be able to achieve high recommendation efficiency, and thus, would not be able to increase their profit. As such, each company has to find the best trade-off between hiding data and training the machine learning algorithm together.Another example comes from the domain of smart grids. Consider a set of households who together would like to form a coalition and trade for good energy deals in the market. In order to maximise their own utility, as well as the social welfare, these households are encouraged to share as much information about their personal preferences with each other as possible. However, this raises a number of privacy issues. Given this, the main objective here is to identify to what extent the homeowners should lower their privacy demands in order to achieve good social welfare (i.e., good collective deals). The abovementioned examples belong to a new and emerging class of optimisation problems called Coopetition (for cooperation in competition), rooted in Game Theory and Bi-Level Optimisation. With the rapid improvements in Data Science, Decentralised Machine Learning (predicted to be the future of Machine Learning), and large-scale ubiquitous intelligent systems (e.g., Internet of Things, smart grids, and smart cities), Coopetition is becoming emergent. However, as it is a very new research field, it still lacks rigorous mathematical models that can capture the complex interconnection between cooperation and competition. As such, during my PhD studies, I would like to investigate the problem of Coopetition and its application to artificial intelligence in more detail. In particular, I aim to achieve the following goals:1. Theoretical foundations: Investigate the mathematical foundations of Coopetition, and propose a thorough theoretical analysis of different models (year 1).2. Algorithmic Contribution: Develop scalable and efficient algorithms to calculate different solution concepts of Coopetition (year 2).3. Application: Apply my theoretical and algorithmic findings to concrete domains of Artificial Intelligence and Machine Learning (year 3).

考虑一个分散的机器学习问题，其中两家竞争对手的在线营销公司旨在共享他们的数据，以便训练推荐算法以最大限度地提高其准确性，因为它们都没有足够大的数据来这样做。他们决定共享的数据越多，他们可以获得的算法性能的改善就越好。然而，由于它们也是竞争对手，彼此共享太多数据也会引起成本（例如，将敏感数据和市场输给竞争对手）。另一方面，如果没有合作，他们就无法实现高推荐效率，从而无法增加利润。因此，每家公司都必须在隐藏数据和训练机器学习算法之间找到最佳平衡。另一个例子来自智能电网领域。考虑一组家庭，他们希望组成一个联盟，并在市场上进行良好的能源交易。为了最大限度地利用自己的效用和社会福利，这些家庭被鼓励尽可能多地相互分享关于他们个人偏好的信息。然而，这引起了一些隐私问题。有鉴于此，这里的主要目标是确定房主应该在多大程度上降低他们的隐私要求，以实现良好的社会福利（即，良好的集体协议）。上述例子属于一类新的和新兴的优化问题，称为合作竞争（竞争中的合作），植根于博弈论和双层优化。随着数据科学、去中心化机器学习（被预测为机器学习的未来）和大规模无处不在的智能系统（例如，物联网、智能电网和智能城市），合作竞争正变得越来越迫切。然而，由于它是一个非常新的研究领域，它仍然缺乏严格的数学模型，可以捕捉合作与竞争之间的复杂联系。因此，在我的博士研究期间，我想更详细地研究合作竞争问题及其在人工智能中的应用。具体而言，我的目标是实现以下目标：1。理论基础：研究合作竞争的数学基础，并提出不同模型的深入理论分析（1年）。2.学术贡献：开发可扩展和有效的算法来计算不同的解决方案概念的合作竞争（第2年）。3.应用：将我的理论和算法发现应用于人工智能和机器学习的具体领域（3年级）。