人工智能培训

搜索

机器学习论文:双人零和游戏中代理理性的大规模学习(Large Scale Learning of Agent Rationality in Two-Player

[复制链接]
imissa 发表于 2019-3-12 11:25:36 | 显示全部楼层 |阅读模式
imissa 2019-3-12 11:25:36 540 0 显示全部楼层
机器学习论文:双人零和游戏中代理理性的大规模学习(Large Scale Learning of Agent Rationality in Two-Player Zero-Sum Games)随着最近在解决大型,零和广泛形式游戏方面的进展,人们越来越关注仅在访问代理动作的情况下推断潜在游戏参数的反问题。虽然最近的工作提供了一个强大的可区分的端到端学习框架,它将游戏手柄嵌入深度学习框架中,允许通过反向传播来学习未知的游戏参数,但当应用于有限理性的人类代理和大规模问题时,该框架面临着重大限制,领先的topoor实用性。在本文中,我们解决了这些限制并提出了适用于更实际环境的框架。首先,在复杂的双人零和游戏中寻求理解人类代理人的合理性,在决策理论中借鉴众所周知的思想,以获得简洁和可解释的代理人行为模型,并得出解决者和梯度的前端到终点学习。其次,为了扩展到大型的真实世界场景,我们提出了一种有效的一阶原始 - 对偶方法,该方法利用了广泛形式游戏的结构,为游戏求解和梯度计算提供了明显更快的计算。在随机生成的游戏中进行测试时,我们会报告比以前的方法更快的数量级。我们还展示了我们的模型在真实世界单人游戏设置和合成数据上的有效性。
With the recent advances in solving large, zero-sum extensive form games,there is a growing interest in the inverse problem of inferring underlying gameparameters given only access to agent actions.Although a recent work providesa powerful differentiable end-to-end learning frameworks which embed a gamesolver within a deep-learning framework, allowing unknown game parameters to belearned via backpropagation, this framework faces significant limitations whenapplied to boundedly rational human agents and large scale problems,leading topoor practicality.In this paper, we address these limitations and propose aframework that is applicable for more practical settings.First, seeking tolearn the rationality of human agents in complex two-player zero-sum games, wedraw upon well-known ideas in decision theory to obtain a concise andinterpretable agent behavior model, and derive solvers and gradients forend-to-end learning.Second, to scale up to large, real-world scenarios, wepropose an efficient first-order primal-dual method which exploits thestructure of extensive-form games, yielding significantly faster computationfor both game solving and gradient computation.When tested on randomlygenerated games, we report speedups of orders of magnitude over previousapproaches.We also demonstrate the effectiveness of our model on bothreal-world one-player settings and synthetic data.机器学习论文:双人零和游戏中代理理性的大规模学习(Large Scale Learning of Agent Rationality in Two-Player Zero-Sum Games) wf8943J7494f5848.jpg
URL地址:https://arxiv.org/abs/1903.04101     ----pdf下载地址:https://arxiv.org/pdf/1903.04101    ----机器学习论文:双人零和游戏中代理理性的大规模学习(Large Scale Learning of Agent Rationality in Two-Player Zero-Sum Games)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

imissa当前离线
新手上路

查看:540 | 回复:0

快速回复 返回顶部 返回列表