Automation and Contemplation for Model Adaptation in Multiagent Interactions

智能体是一种计算机系统，它根据来自环境的感官输入智能地行动。Agent技术在许多实际应用中被证明是有效和可靠的解决方案，并将继续在现代社会中发挥重要作用。例如，eBay买家代理为电子市场中的人们推荐好的交易。由自动驾驶代理操作的谷歌自动驾驶汽车已经成功地在道路上行驶了数千英里。由智能软件代理控制的智能电表有助于优化家庭能源消耗。在许多这样的应用中，一个自治的代理（即主体代理）被期望通过预测在一个共同的环境中的其他代理的行为来做出理性的决策。决策质量依赖于建立其他代理的决策模型，然后求解模型以了解其他代理在环境中的行为。当主体代理的模型部署在真实世界的应用程序中时，它可能会失败，因为主体代理可能会收到其他代理引起的意外观察。因此，挑战是关于预测其他代理的行为和解释模型失败，以适应主体代理的模型成功的互动。该项目的目标是通过自动化其他代理的模型构建和修改自己的决策模型时，模型在执行失败，以提高主体代理的适应性。这个项目将提出可扩展的学习算法，以建立决策模型的其他代理人的历史数据的代理人的相互作用。该算法还将促进在新的问题域中的模型构建，该问题域在实践中可能更大且更不确定。为了解释主体智能体的决策模型的失败，本项目将寻找一种新的推理技术，以确定失败背后的最可能的原因，并相应地修改模型，使主体智能体的决策可以适应其他智能体的行为，在他们的实时交互。该项目将在一个工具包中实现所有拟议的技术，并进行全面的测试，以评估该工具包的实际效用。个性化学习和智能计算机游戏AI引擎开发的实际应用将通过我们的行业合作伙伴进行扩展。这项研究的更广泛的影响将是使个体智能体能够在复杂的多智能体环境中理性地行动。这是将自主代理技术整合到社会中的关键一步，将支持人类执行灾难响应、能源分配和安全操作等任务。

An agent is a computer system that acts intelligently given its sensory input from the environment. Agent technologies have proved to be effective and reliable solutions in many practical applications and will continue to play a major role in modern society. For example, the eBay buyer agent recommends good deals for people in an e-market. The Google self-driving car operated by an autonomous agent has successfully navigated thousands of miles on the road. A smart meter controlled by an intelligent software agent helps optimize energy consumption for a household. In many such applications, an autonomous agent (namely a subject agent) is expected to make a rational decision by predicting behaviors of other agents in a common environment. The decision quality relies on building decision models of the other agents and then solving the models to understand how the other agents will behave in the environment. When the subject agent's model is deployed in a real-world application, it may fail since the subject agent may receive unexpected observations incurred by other agents. Hence the challenge is about the prediction of other agents' behavior and the interpretation of model failure so as to adapt the subject agent's model for successful interactions. The goal of this project is to improve the subject agent's adaptation by automating the model construction of other agents and revising its own decision model when the model fails in the execution. This project will propose scalable learning algorithms to build decision models of other agents upon historical data of agents' interactions. The algorithms will also facilitate the model construction in a new problem domain that will be likely larger and more uncertain in practice. To interpret failures of the subject agent's decision model, this project will search for a novel reasoning technique to identify the most probable reasons behind the failures, and accordingly revise the model so that the subject agent's decisions can be adapted to the other agents' behaviors in their real-time interactions. This project will implement all the proposed techniques in a toolkit and conduct comprehensive tests to evaluate practical utilities of the toolkit. Real-world applications on personalized learning and intelligent computer game AI engine development will be extended through our industrial collaborators. The broader impact of this research will be to enable individual agents to act rationally in complex multiagent environments. This is a crucial step toward the integration of autonomous agent technology within society that will support humans in tasks such as disaster response, energy distribution and security operation.