When Does Communication Learning Need Hierarchical Multi-Agent Deep Reinforcement Learning

Abstract

Multi-agent systems need to communicate to coordinate a shared task. We show that a recurrent neural network (RNN) can learn a communication protocol for coordination, even if the actions to coordinate are performed steps after the communication phase. We show that a separation of tasks with different temporal scale is necessary for successful learning. We contribute a hierarchical deep reinforcement learning model for multi-agent systems that separates the communication and coordination task from the action picking through a hierarchical policy. We further on show, that a separation of concerns in communication is beneficial but not necessary. As a testbed, we propose the Dungeon Lever Game and we extend the Differentiable Inter-Agent Learning (DIAL) framework. We present and compare results from different model variations on the Dungeon Lever Game.     

异构数据集成方法研究与实现

为解决异构数据集成的复杂性问题，文章提出了一种异构数据同步的新方法——半集成式中间件法。该方法成功地应用在重庆市社会保障综合信息系统中，具有易于实现、可扩展性、跨平台性和效率高等特点。     

由AEO_3硫酸化合成AES的研究

以氨基磺酸和氯磺酸作硫酸化剂,由AEO_3合成了AES,讨论了工艺条件对反应过程和产品质量的影响。     

钢包钢水保温覆盖剂的开发

研制了钢包钢水新型保温覆盖剂，主要成分为（%）；15-25C，20-35SiO2,10-20Al2O3,2-5Fe2O3,15-25CaO,2-5MgO，新覆盖剂铺展性与保温性能良好。经转炉钢水160t钢包生产使用，统计数据表明，镇静15min后钢水降温速度由原碳化稻壳覆盖剂1.33℃/min降低到0.93℃/min。     

A general formula for the evaluation of thermodynamic and aerodynamic losses in nucleating steam flow

Near saturation steam undergoing rapid expansion, with homogeneous nucleation of water droplets, is numerically studied in a series of converging/diverging nozzles with and without shocks. To understand loss mechanisms in such flows a numerical model is presented to calculate thermodynamic losses, which is further used to quantify associated total aerodynamic losses. For the converging/diverging nozzle configuration, the model shows that the overall thermodynamic loss is only mildly influenced by increasing shock strength, while the aerodynamic losses follow that of the single phase flow, and are of the same magnitude as the thermodynamic loss only in the case of very weak shocks. The thermodynamic losses can be attributed to two influences, the homogeneous nucleation event, and the post-shock thermal oscillations in the two-phase system. The calculations rely on a new two-phase CFD model, previously reported, for non-equilibrium phase change with droplet nucleation applicable to general 3D flow configurations.     

基于CORBA和多Agent的模具敏捷制造的工作流管理

介绍了CORBA的工作原理，分析了模具敏捷制造的特点，提出了基于CORBA的模具敏捷制造工作流管理模型，并且将多Agent技术应用到该模型之中，使其具有清晰的结构，良好的开放性、集成性和自适应性。     

联合法制备VBL萤光增白剂－砍两头新工艺

将ＤＳＤ酸生产中还原反应并经中和过滤后的清液直接用于ＶＢＬ萤光增白剂生产，获得满意的效果。     

基于ObjectAgent的小卫星星务系统研究与仿真

针对小卫星自主管理与自主决策的要求,简单介绍了多Agent技术和ObjectAgent开发环境,并以某小卫星为例,研究并设计了基于ObjectAgent的小卫星星务系统,给出了所涉及的Agent的自主决策、任务调度、信息交互等一系列问题的解决方案,最后,利用多台PC机组网进行了仿真,达到了预期的实验效果。     

一种新型硫酸亚铁单液法调剖剂

本文提出用尿素作为ｐⅡ调节剂，加快硫酸亚铁单液法调剖剂中亚铁离子在地层温度下的沉淀，增大沉淀率，提高封堵效能。岩芯流动试验结果表明，１５％ＦｅＳＯ４．７ＩＩ ２Ｏ／５％尿素调剖剂体系的封堵能力比不含尿素的体系增大近一倍。     

Agens Faber: Toward a Theory of Artefacts for MAS

Human intelligence has evolved along with the use of more and more sophisticated tools, allowing Homo Faber (from Homo Habilis to Homo Sapiens Sapiens) to cope with environment changes, as well as to adapt the environment to his needs. Analogously, in this seminal paper we introduce the notion of Agens Faber, conveying the idea that agent intelligence should not be considered as separated by the agent ability to perceive and affect the environment—and so, that agent intelligence is strictly related to the artefacts that enable, mediate and govern any agent (intelligent) activity.

Along this line, we first discuss the notion of artefact for MAS in general, then we try to devise out the admissible / required / desirable features of an artefact for MAS. We elaborate on the many sorts of possible relations between agents and artefacts, focusing in particular on the issue of the rational exploitation of artefacts, and also rough out a possible taxonomy of artefacts for MAS.