CIF: Small: Strategic Communication: Concepts, Methods, and Applications

我们被智能代理网络所包围，这些智能代理通过通信渠道相互交换信息，就像物联网和自主系统一样。本项目考虑了当代理的目标不一致或冲突时，与这些智能和战略代理之间的沟通有关的问题，称为“战略沟通”。将已知的战略通信模型扩展到涉及信道噪声和数据压缩的现实通信设置仍然是一个开放的挑战。这个项目的目标是为这种现实场景开发战略沟通的数学模型。这项研究的直接应用涉及可信和透明的机器学习（ML）。机器学习越来越多地被用于做出影响人类的决策的系统，包括设定物品价格、信用评分和工作申请、过滤新闻和社交媒体更新、推荐路线和地点，以及控制智能家居和自动驾驶汽车。当机器学习算法做出这些决定时，要求它们的透明度是很自然的，另一方面，这使得它们容易受到操纵和算法偏见的影响。设计对操纵和偏见具有鲁棒性的高效透明的ML算法是一项艰巨的挑战。该项目将通过强大的战略通信模型在ML算法开发中明确考虑这些可能性，从而解决这一挑战。这项研究将解决有关战略沟通的基本问题。这一新兴研究领域需要重新审视经典信息理论的关键成果，并在分析和优化方面提出了重大挑战，需要来自多个学科的方法和工具。研究结果有望成为理解博弈论和经济学与信息理论、通信和压缩之间相互作用的重要一步。这个项目的具体目标分为四组。第一组目标研究了非编码通信设置中的最佳策略，特别关注使用功能空间优化工具的非高斯源和通道。第二组目标涉及机器学习中的透明度问题。基于非编码场景中的战略通信模型，该项目包括该领域的两个研究方向：各种机器学习算法透明度成本的量化和操作感知机器学习算法的设计。第三组目标分析了压缩在战略通信场景中的作用，包括基本限制的表征和计算，以及实用战略数据压缩方法的发展。最后一组目标探讨网络化战略传播场景。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知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

We are surrounded by networks of smart agents that exchange information with each other through communication channels, as in the case of Internet of Things and autonomous systems. This project considers problems related to communication between these smart and strategic agents referred to as 'strategic communication', when the agents have misaligned or conflicting objectives. Extensions of known strategic communication models to realistic communication settings that involve channel noise and data compression remains an open challenge. The goal of this project is to develop mathematical models of strategic communication for such realistic scenarios. An immediate application of this research pertains to trustworthy and transparent machine learning (ML). ML is increasingly being used in systems that make decisions which affect people, including setting prices for items, scoring credit and job applications, filtering news and social media updates, recommending routes and places, and controlling smart homes and autonomous cars. As ML algorithms make these decisions, it is natural to ask for their transparency, which, on the other hand, makes them vulnerable to manipulation and algorithmic bias. Designing efficient and transparent ML algorithms that are robust to manipulation and bias is a difficult challenge. This project will address this challenge by explicitly taking such possibilities into account in the ML algorithm development through robust strategic communication models. This research will address fundamental questions regarding strategic communication. This emerging research field requires revisiting key results in classical information theory and poses significant challenges, in terms of both analysis and optimization, requiring approaches and tools from multiple disciplines. The outcomes of the research are expected to constitute an essential step in understanding the interplay of game theory and economics with information theory, communications, and compression. Specific goals of this project are categorized into four groups. The first set of goals investigates optimal strategies in non-coded communication settings, with a particular focus on non-Gaussian sources and channels using tools from optimization in function spaces. The second set of goals concerns the transparency issues in ML. Building on the strategic communication models in non-coded scenarios, the project includes two research directions in this area: the quantification of the cost of transparency in various ML algorithms and the design of manipulation-aware ML algorithms. The third set of goals analyzes the role of compression in strategic communication scenarios, including the characterization and computation of fundamental limits, and the development of practical strategic data compression methods. The final set of goals explores networked strategic communication scenarios.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.