Evolution of Communication in Multi-Agent Systems

这个项目将探索通信和解决问题的方式在自主智能体的组中相互作用，例如移动机器人团队，并研究在这些组中发展任务特定语言的方法。其目标是提出新的算法和协议，利用进化算法的力量将问题解决和交流结合起来。更具体地说，PI将试图证明，在其约束可以建模为进化压力的任务环境中，一组自治代理可能会进化出一种特定于任务的通信系统和简单的语言，从而提高它们的整体效率。此外，在监视任务或嘈杂的通信网络等惩罚通信努力的环境中，仅包括完成任务所需的言语行为的特定任务通信系统将最大限度地减少代理间的通信，从而显著提高群体问题解决能力。这个多方面的问题将通过将其框定在有据可查的捕食者/猎物领域来解决。在工程领域，捕食者被视为执行搜索和救援或分布式空间探索等任务的多机器人群体的代表。这一环境将扩展到包括对具有不同感知和执行能力的广义、不同种类的机器人的组的分析，这些机器人必须有效地协调以实现组目标。在这种任务中，环境的确切性质可能不是先验的。因此，将使用一种模拟工具，允许通信能力演变，以便最好地将代理的能力与未知或部分已知的任务环境相匹配。模拟是模块化的，允许包括或排除各种选择力量，以便隔离使有效沟通得以发展的因素。PI将开发一种硬件试验台，可供许多学生用于独立研究项目。这张试验床也将是K-12极端外展工作的亮点。

This project will explore the way that communication and problem solving interact in groups of autonomous agents, such as mobile robot teams, and to examine methods for evolving task-specific language in these groups. The goal is to come up with new algorithms and protocols that exploit the power of evolutionary algorithms to combine problem solving and communication. More specifically, the PI will try to demonstrate that in task environments whose constraints can be modeled as evolutionary pressures, a group of autonomous agents may evolve a task-specific communication system and simple language that will increase their overall effectiveness. Moreover in environments that penalize communication efforts such as surveillance tasks or noisy communication networks, a task-specific communication system including only those speech acts found necessary to complete the task will minimize inter-agent communications, yielding significant group problem-solving improvement. This multi-faceted problem will be approached by framing it within the well-documented Predator/Prey domain. In the realm of engineering, the predators are seen as representative of a multi-robot group performing a task such as search and rescue or distributed space exploration. This environment will be extended to include the analysis of groups of generalized, heterogeneous robots with different sensing and actuation abilities, which must coordinate efficiently to achieve the group goal. The exact nature of the environment in such tasks may not be known a priori. Hence, a simulation tool will be used that allows communication abilities to evolve in order to best match the agent's abilities to an unknown or partially known task environment. The simulation is modular and allows for the inclusion or exclusion of various selection forces in order to isolate the factors that enable effective communication to develop. The PI will develop a hardware testbed that can be used by many students for independent research projects. This testbed will also be the highlight of an extreme K-12 outreach effort.