权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

Cooperative and Non-Cooperative artificial intelligence (AI)

合作型和非合作型人工智能 (AI)

基本信息

批准号：
2574825
负责人：
金额：
--
依托单位：
University College London
依托单位国家：
英国
项目类别：
Studentship
财政年份：
2021
资助国家：
英国
起止时间：
2021 至无数据
项目状态：
未结题

来源：
https://gtr.ukri.org/projects?ref=studentship-2574825
关键词：
Cooperative Non artificial intelligence AI

项目摘要

For my project I intend to study how AI agents can learn to cooperate with one another and take advantage of one another in general sum games. The games will be social dilemmas, meaning the players of these games can make a personal choice to increase their reward, at the detriment to the welfare of the other players. To begin my study of this topic I first intend to do an information theoretical study of various deterministic strategies used to play the prisoners dilemma (a simple social dilemma), as outlined by Axelrod, as well as including simple naïve learning agents. This should allow me to outline metrics that can be used to classify strategies, by asking questions such as, does this strategy rely on a transfer of information from its opponent? If it does, does it use this information to reciprocate behaviours? Once these strategies have been classified using information theoretic metrics it should be possible to detect agents that change strategy, including agents who try go from cooperating with their opponent to then trying to exploit them. Allowing us to prepare agents who can detect and react to deceptive behaviours from their opponent. Finally, we expect it to be possible to use these metrics to try and inform the design of deep reinforcement learning agents that can learn to take advantage of their opponent.

在我的项目中，我打算研究人工智能智能如何学会相互合作，并在一般和游戏中相互利用。这些游戏将是社会困境，这意味着这些游戏的玩家可以做出个人选择来增加他们的奖励，而损害其他玩家的福利。为了开始我对这个主题的研究，我首先打算对阿克塞尔罗德概述的用于解决囚犯困境（一种简单的社会困境）的各种确定性策略进行信息理论研究，以及包括简单的天真学习代理。这应该允许我概述可用于对策略进行分类的指标，通过提出诸如，该策略是否依赖于来自对手的信息传输之类的问题？如果是，它会利用这些信息来回报行为吗？一旦这些策略已经使用信息理论度量分类，就应该有可能检测到改变策略的代理，包括试图从与对手合作到试图利用他们的代理。使我们能够准备代理人谁可以检测和反应，从他们的对手的欺骗行为。最后，我们希望有可能使用这些指标来尝试和通知深度强化学习代理的设计，这些代理可以学习利用对手。