基于自主学习的Ad hoc Agent序贯决策研究

Multi-agent decision making techniques always assume cooperative agents that can resolve pre-defined tasks through communication and coordination. The techniques however are not applicable for solving decision problems with competitive agents. It is a challenge to develop an adaptive agent, namely Ad hoc agent, that can construct and solve decision problems in an environment commonly shared by other agents of unknown relationships. This project will solve sequential decision making problems involving Ad hoc agents from individual agent perspective. A subject agent will learn behavior of other ad hoc agents by adapting machine learning techniques, and accordingly update its own decision models. This project will extend learning algorithms for constructing behavioral model of a single agent to learn behavioral patterns of a population of other agents. Based on the scenario of unmanned aerial vehicle, this project will build a platform for simulating interactions, performing learning and conducting evaluation for ad hoc agents. In summary, this project will develop a new type of Ad hoc agent that can actively explore the environment with other unknown agents. The research outcomes will facilitate applications of multi-agent technologies in complex problem domains, and provide theoretical guarantees and practical guidelines.

多智能体（Agent）决策技术的研究常假设智能体之间通过通信与协调来完成既定任务。该假设不适用于具有竞争关系的多智能体系统。因此，在未知决策环境下，开发具有自适应能力的智能体，即Ad hoc Agent，是多智能体研究领域极具挑战的新兴问题。本项目将提出一个基于个体智能体自主学习与决策的新框架，以构造并求解多Ad hoc Agent序贯决策问题。其主要研究内容包括：通过机器学习方法，使Ad hoc Agent能从交互数据中自主构造出准确刻画其他智能体行为特征的模型，并更新自身的决策模型；在此基础上，将针对个体智能体行为模型的学习算法，推广到学习群体智能体抽象行为中；最终搭建一个以无人驾驶飞机仿真为背景的Ad hoc Agent仿真平台。本项目期望构造能自主发掘并合理应对陌生智能体行为的新型Ad hoc Agent，为将多智能体技术应用于更加复杂多变的现实场景中，提供理论依据与实践指导。

多智能体（Agent）决策技术的研究常假设智能体之间通过通信与协调来完成既定任务。该假设不适用于具有竞争关系的多智能体系统。因此，在未知决策环境下，开发具有自适应能力的智能体，即Ad hoc Agent，是多智能体研究领域的极具挑战的新兴问题。本项目提出了一个基于个体智能体自主学习与决策的新框架，以求解多Ad hoc Agent序贯决策问题。其主要研究内容包括：通过机器学习方法，使Ad hoc Agent能从交互数据中自主构造出准确刻画其他智能体行为特征的模型，并采用模型检测技术对智能体行为进行分析；结合博弈论，研究了贝叶斯智能体的类型对决策过程的影响。此外，本项目将针对个体智能体的决策方法，推广到群体智能体中，实现了千个智能体的交互。本项目除了搭建一个以无人驾驶飞机仿真为背景的Ad hoc Agent交互、学习、验证平台，还以多人在线手机游戏为真实测试载体，验证了智能体行为对人类玩家行为的理性应对方式。