arXiv - CS - Multiagent Systems最新文献_第2页

On the limits of agency in agent-based models 基于代理的模型中代理的局限性

arXiv - CS - Multiagent Systems Pub Date : 2024-09-14 DOI: arxiv-2409.10568

Ayush Chopra, Shashank Kumar, Nurullah Giray-Kuru, Ramesh Raskar, Arnau Quera-Bofarull

{"title":"On the limits of agency in agent-based models","authors":"Ayush Chopra, Shashank Kumar, Nurullah Giray-Kuru, Ramesh Raskar, Arnau Quera-Bofarull","doi":"arxiv-2409.10568","DOIUrl":"https://doi.org/arxiv-2409.10568","url":null,"abstract":"Agent-based modeling (ABM) seeks to understand the behavior of complex\u0000systems by simulating a collection of agents that act and interact within an\u0000environment. Their practical utility requires capturing realistic environment\u0000dynamics and adaptive agent behavior while efficiently simulating million-size\u0000populations. Recent advancements in large language models (LLMs) present an\u0000opportunity to enhance ABMs by using LLMs as agents with further potential to\u0000capture adaptive behavior. However, the computational infeasibility of using\u0000LLMs for large populations has hindered their widespread adoption. In this\u0000paper, we introduce AgentTorch -- a framework that scales ABMs to millions of\u0000agents while capturing high-resolution agent behavior using LLMs. We benchmark\u0000the utility of LLMs as ABM agents, exploring the trade-off between simulation\u0000scale and individual agency. Using the COVID-19 pandemic as a case study, we\u0000demonstrate how AgentTorch can simulate 8.4 million agents representing New\u0000York City, capturing the impact of isolation and employment behavior on health\u0000and economic outcomes. We compare the performance of different agent\u0000architectures based on heuristic and LLM agents in predicting disease waves and\u0000unemployment rates. Furthermore, we showcase AgentTorch's capabilities for\u0000retrospective, counterfactual, and prospective analyses, highlighting how\u0000adaptive agent behavior can help overcome the limitations of historical data in\u0000policy design. AgentTorch is an open-source project actively being used for\u0000policy-making and scientific discovery around the world. The framework is\u0000available here: github.com/AgentTorch/AgentTorch.","PeriodicalId":501315,"journal":{"name":"arXiv - CS - Multiagent Systems","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-09-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142258971","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Swarm Algorithms for Dynamic Task Allocation in Unknown Environments 用于未知环境中动态任务分配的蜂群算法

arXiv - CS - Multiagent Systems Pub Date : 2024-09-14 DOI: arxiv-2409.09550

Adithya Balachandran, Noble Harasha, Nancy Lynch

{"title":"Swarm Algorithms for Dynamic Task Allocation in Unknown Environments","authors":"Adithya Balachandran, Noble Harasha, Nancy Lynch","doi":"arxiv-2409.09550","DOIUrl":"https://doi.org/arxiv-2409.09550","url":null,"abstract":"Robot swarms, systems of many robots that operate in a distributed fashion,\u0000have many applications in areas such as search-and-rescue, natural disaster\u0000response, and self-assembly. Several of these applications can be abstracted to\u0000the general problem of task allocation in an environment, in which robots must\u0000assign themselves to and complete tasks. While several algorithms for task\u0000allocation have been proposed, most of them assume either prior knowledge of\u0000task locations or a static set of tasks. Operating under a discrete general\u0000model where tasks dynamically appear in unknown locations, we present three new\u0000swarm algorithms for task allocation. We demonstrate that when tasks appear\u0000slowly, our variant of a distributed algorithm based on propagating task\u0000information completes tasks more efficiently than a Levy random walk algorithm,\u0000which is a strategy used by many organisms in nature to efficiently search an\u0000environment. We also propose a division of labor algorithm where some agents\u0000are using our algorithm based on propagating task information while the\u0000remaining agents are using the Levy random walk algorithm. Finally, we\u0000introduce a hybrid algorithm where each agent dynamically switches between\u0000using propagated task information and following a Levy random walk. We show\u0000that our division of labor and hybrid algorithms can perform better than both\u0000our algorithm based on propagated task information and the Levy walk algorithm,\u0000especially at low and medium task rates. When tasks appear fast, we observe the\u0000Levy random walk strategy performs as well or better when compared to these\u0000novel approaches. Our work demonstrates the relative performance of these\u0000algorithms on a variety of task rates and also provide insight into optimizing\u0000our algorithms based on environment parameters.","PeriodicalId":501315,"journal":{"name":"arXiv - CS - Multiagent Systems","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-09-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142258975","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Mutual Theory of Mind in Human-AI Collaboration: An Empirical Study with LLM-driven AI Agents in a Real-time Shared Workspace Task 人机协作中的相互心智理论：在实时共享工作区任务中使用 LLM 驱动的人工智能代理的实证研究

arXiv - CS - Multiagent Systems Pub Date : 2024-09-13 DOI: arxiv-2409.08811

Shao Zhang, Xihuai Wang, Wenhao Zhang, Yongshan Chen, Landi Gao, Dakuo Wang, Weinan Zhang, Xinbing Wang, Ying Wen

{"title":"Mutual Theory of Mind in Human-AI Collaboration: An Empirical Study with LLM-driven AI Agents in a Real-time Shared Workspace Task","authors":"Shao Zhang, Xihuai Wang, Wenhao Zhang, Yongshan Chen, Landi Gao, Dakuo Wang, Weinan Zhang, Xinbing Wang, Ying Wen","doi":"arxiv-2409.08811","DOIUrl":"https://doi.org/arxiv-2409.08811","url":null,"abstract":"Theory of Mind (ToM) significantly impacts human collaboration and\u0000communication as a crucial capability to understand others. When AI agents with\u0000ToM capability collaborate with humans, Mutual Theory of Mind (MToM) arises in\u0000such human-AI teams (HATs). The MToM process, which involves interactive\u0000communication and ToM-based strategy adjustment, affects the team's performance\u0000and collaboration process. To explore the MToM process, we conducted a\u0000mixed-design experiment using a large language model-driven AI agent with ToM\u0000and communication modules in a real-time shared-workspace task. We find that\u0000the agent's ToM capability does not significantly impact team performance but\u0000enhances human understanding of the agent and the feeling of being understood.\u0000Most participants in our study believe verbal communication increases human\u0000burden, and the results show that bidirectional communication leads to lower\u0000HAT performance. We discuss the results' implications for designing AI agents\u0000that collaborate with humans in real-time shared workspace tasks.","PeriodicalId":501315,"journal":{"name":"arXiv - CS - Multiagent Systems","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-09-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142258631","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

CollaMamba: Efficient Collaborative Perception with Cross-Agent Spatial-Temporal State Space Model CollaMamba：利用跨代理时空状态模型实现高效协作感知

arXiv - CS - Multiagent Systems Pub Date : 2024-09-12 DOI: arxiv-2409.07714

Yang Li, Quan Yuan, Guiyang Luo, Xiaoyuan Fu, Xuanhan Zhu, Yujia Yang, Rui Pan, Jinglin Li

{"title":"CollaMamba: Efficient Collaborative Perception with Cross-Agent Spatial-Temporal State Space Model","authors":"Yang Li, Quan Yuan, Guiyang Luo, Xiaoyuan Fu, Xuanhan Zhu, Yujia Yang, Rui Pan, Jinglin Li","doi":"arxiv-2409.07714","DOIUrl":"https://doi.org/arxiv-2409.07714","url":null,"abstract":"By sharing complementary perceptual information, multi-agent collaborative\u0000perception fosters a deeper understanding of the environment. Recent studies on\u0000collaborative perception mostly utilize CNNs or Transformers to learn feature\u0000representation and fusion in the spatial dimension, which struggle to handle\u0000long-range spatial-temporal features under limited computing and communication\u0000resources. Holistically modeling the dependencies over extensive spatial areas\u0000and extended temporal frames is crucial to enhancing feature quality. To this\u0000end, we propose a resource efficient cross-agent spatial-temporal collaborative\u0000state space model (SSM), named CollaMamba. Initially, we construct a\u0000foundational backbone network based on spatial SSM. This backbone adeptly\u0000captures positional causal dependencies from both single-agent and cross-agent\u0000views, yielding compact and comprehensive intermediate features while\u0000maintaining linear complexity. Furthermore, we devise a history-aware feature\u0000boosting module based on temporal SSM, extracting contextual cues from extended\u0000historical frames to refine vague features while preserving low overhead.\u0000Extensive experiments across several datasets demonstrate that CollaMamba\u0000outperforms state-of-the-art methods, achieving higher model accuracy while\u0000reducing computational and communication overhead by up to 71.9% and 1/64,\u0000respectively. This work pioneers the exploration of the Mamba's potential in\u0000collaborative perception. The source code will be made available.","PeriodicalId":501315,"journal":{"name":"arXiv - CS - Multiagent Systems","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142190479","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Self-Supervised Inference of Agents in Trustless Environments 无信任环境中的代理自监督推理

arXiv - CS - Multiagent Systems Pub Date : 2024-09-12 DOI: arxiv-2409.08386

Vladyslav Larin, Ivan Nikitin, Alexander Firsov

引用次数: 0

Simultaneous Topology Estimation and Synchronization of Dynamical Networks with Time-varying Topology 时变拓扑动态网络的同步拓扑估计与同步

arXiv - CS - Multiagent Systems Pub Date : 2024-09-12 DOI: arxiv-2409.08404

Nana Wang, Esteban Restrepo, Dimos V. Dimarogonas

引用次数: 0

Leveraging Unstructured Text Data for Federated Instruction Tuning of Large Language Models 利用非结构化文本数据对大型语言模型进行联合教学调整

arXiv - CS - Multiagent Systems Pub Date : 2024-09-11 DOI: arxiv-2409.07136

Rui Ye, Rui Ge, Yuchi Fengting, Jingyi Chai, Yanfeng Wang, Siheng Chen

{"title":"Leveraging Unstructured Text Data for Federated Instruction Tuning of Large Language Models","authors":"Rui Ye, Rui Ge, Yuchi Fengting, Jingyi Chai, Yanfeng Wang, Siheng Chen","doi":"arxiv-2409.07136","DOIUrl":"https://doi.org/arxiv-2409.07136","url":null,"abstract":"Federated instruction tuning enables multiple clients to collaboratively\u0000fine-tune a shared large language model (LLM) that can follow humans'\u0000instructions without directly sharing raw data. However, existing literature\u0000impractically requires that all the clients readily hold instruction-tuning\u0000data (i.e., structured instruction-response pairs), which necessitates massive\u0000human annotations since clients' data is usually unstructured text instead.\u0000Addressing this, we propose a novel and flexible framework FedIT-U2S, which can\u0000automatically transform unstructured corpus into structured data for federated\u0000instruction tuning. FedIT-U2S consists two key steps: (1) few-shot\u0000instruction-tuning data generation, where each unstructured data piece together\u0000with several examples is combined to prompt an LLM in generating an\u0000instruction-response pair. To further enhance the flexibility, a\u0000retrieval-based example selection technique is proposed, where the examples are\u0000automatically selected based on the relatedness between the client's data piece\u0000and example pool, bypassing the need of determining examples in advance. (2) A\u0000typical federated instruction tuning process based on the generated data.\u0000Overall, FedIT-U2S can be applied to diverse scenarios as long as the client\u0000holds valuable text corpus, broadening the application scope of federated\u0000instruction tuning. We conduct a series of experiments on three domains\u0000(medicine, knowledge, and math), showing that our proposed FedIT-U2S can\u0000consistently and significantly brings improvement over the base LLM.","PeriodicalId":501315,"journal":{"name":"arXiv - CS - Multiagent Systems","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-09-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142190482","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

DCMAC: Demand-aware Customized Multi-Agent Communication via Upper Bound Training DCMAC：通过上限训练实现需求感知的定制多代理通信

arXiv - CS - Multiagent Systems Pub Date : 2024-09-11 DOI: arxiv-2409.07127

Dongkun Huo, Huateng Zhang, Yixue Hao, Yuanlin Ye, Long Hu, Rui Wang, Min Chen

{"title":"DCMAC: Demand-aware Customized Multi-Agent Communication via Upper Bound Training","authors":"Dongkun Huo, Huateng Zhang, Yixue Hao, Yuanlin Ye, Long Hu, Rui Wang, Min Chen","doi":"arxiv-2409.07127","DOIUrl":"https://doi.org/arxiv-2409.07127","url":null,"abstract":"Efficient communication can enhance the overall performance of collaborative\u0000multi-agent reinforcement learning. A common approach is to share observations\u0000through full communication, leading to significant communication overhead.\u0000Existing work attempts to perceive the global state by conducting teammate\u0000model based on local information. However, they ignore that the uncertainty\u0000generated by prediction may lead to difficult training. To address this\u0000problem, we propose a Demand-aware Customized Multi-Agent Communication (DCMAC)\u0000protocol, which use an upper bound training to obtain the ideal policy. By\u0000utilizing the demand parsing module, agent can interpret the gain of sending\u0000local message on teammate, and generate customized messages via compute the\u0000correlation between demands and local observation using cross-attention\u0000mechanism. Moreover, our method can adapt to the communication resources of\u0000agents and accelerate the training progress by appropriating the ideal policy\u0000which is trained with joint observation. Experimental results reveal that DCMAC\u0000significantly outperforms the baseline algorithms in both unconstrained and\u0000communication constrained scenarios.","PeriodicalId":501315,"journal":{"name":"arXiv - CS - Multiagent Systems","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-09-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142190499","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Can Agents Spontaneously Form a Society? Introducing a Novel Architecture for Generative Multi-Agents to Elicit Social Emergence 代理能自发形成社会吗？引入新的多代理生成架构以激发社会涌现

arXiv - CS - Multiagent Systems Pub Date : 2024-09-10 DOI: arxiv-2409.06750

H. Zhang, J. Yin, M. Jiang, C. Su

{"title":"Can Agents Spontaneously Form a Society? Introducing a Novel Architecture for Generative Multi-Agents to Elicit Social Emergence","authors":"H. Zhang, J. Yin, M. Jiang, C. Su","doi":"arxiv-2409.06750","DOIUrl":"https://doi.org/arxiv-2409.06750","url":null,"abstract":"Generative agents have demonstrated impressive capabilities in specific\u0000tasks, but most of these frameworks focus on independent tasks and lack\u0000attention to social interactions. We introduce a generative agent architecture\u0000called ITCMA-S, which includes a basic framework for individual agents and a\u0000framework called LTRHA that supports social interactions among multi-agents.\u0000This architecture enables agents to identify and filter out behaviors that are\u0000detrimental to social interactions, guiding them to choose more favorable\u0000actions. We designed a sandbox environment to simulate the natural evolution of\u0000social relationships among multiple identity-less agents for experimental\u0000evaluation. The results showed that ITCMA-S performed well on multiple\u0000evaluation indicators, demonstrating its ability to actively explore the\u0000environment, recognize new agents, and acquire new information through\u0000continuous actions and dialogue. Observations show that as agents establish\u0000connections with each other, they spontaneously form cliques with internal\u0000hierarchies around a selected leader and organize collective activities.","PeriodicalId":501315,"journal":{"name":"arXiv - CS - Multiagent Systems","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-09-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142190481","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A Quality Diversity Approach to Automatically Generate Multi-Agent Path Finding Benchmark Maps 自动生成多代理路径查找基准图的质量多样性方法

arXiv - CS - Multiagent Systems Pub Date : 2024-09-10 DOI: arxiv-2409.06888

Cheng Qian, Yulun Zhang, Varun Bhatt, Matthew Christopher Fontaine, Stefanos Nikolaidis, Jiaoyang Li

{"title":"A Quality Diversity Approach to Automatically Generate Multi-Agent Path Finding Benchmark Maps","authors":"Cheng Qian, Yulun Zhang, Varun Bhatt, Matthew Christopher Fontaine, Stefanos Nikolaidis, Jiaoyang Li","doi":"arxiv-2409.06888","DOIUrl":"https://doi.org/arxiv-2409.06888","url":null,"abstract":"We use the Quality Diversity (QD) algorithm with Neural Cellular Automata\u0000(NCA) to generate benchmark maps for Multi-Agent Path Finding (MAPF)\u0000algorithms. Previously, MAPF algorithms are tested using fixed, human-designed\u0000benchmark maps. However, such fixed benchmark maps have several problems.\u0000First, these maps may not cover all the potential failure scenarios for the\u0000algorithms. Second, when comparing different algorithms, fixed benchmark maps\u0000may introduce bias leading to unfair comparisons between algorithms. In this\u0000work, we take advantage of the QD algorithm and NCA with different objectives\u0000and diversity measures to generate maps with patterns to comprehensively\u0000understand the performance of MAPF algorithms and be able to make fair\u0000comparisons between two MAPF algorithms to provide further information on the\u0000selection between two algorithms. Empirically, we employ this technique to\u0000generate diverse benchmark maps to evaluate and compare the behavior of\u0000different types of MAPF algorithms such as bounded-suboptimal algorithms,\u0000suboptimal algorithms, and reinforcement-learning-based algorithms. Through\u0000both single-planner experiments and comparisons between algorithms, we identify\u0000patterns where each algorithm excels and detect disparities in runtime or\u0000success rates between different algorithms.","PeriodicalId":501315,"journal":{"name":"arXiv - CS - Multiagent Systems","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-09-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142190480","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0