人工智能辅助决策，是实现水下无人系统集群作战应用智能化所面临的关键问题。在实际作战应用中，水下无人系统集群存在装备异构性、约束动态性、任务不确定性等问题。传统的人工智能方法难以解决状态及约束要素动态变化所导致的模型不确定性问题。图神经网络技术是基于认知科学的连接主义人工智能方法——关系型强化学习的一种。通过构建决策图，用决策图的顶点表示无人系统集群智能决策状态及约束要素属性，用决策图的边表示各决策要素之间的逻辑推理关系属性，通过强化学习方法训练整个决策图的顶点属性、边属性以及决策图全局属性，从而通过提取决策图的属性信息实现水下无人系统的人工智能辅助决策。本文旨在探究将图神经网络技术应用在水下无人系统智能辅助决策中的可行性，对水下无人系统智能辅助决策进行图神经网络建模，构建智能辅助决策推理算法伪代码，研究基于图神经网络技术的水下无人系统智能决策的技术实现。
Artificial intelligent decision support is a key problem to realize intelligent application of underwater unmanned system cluster operation. In practical operational applications, underwater unmanned system clusters are confronted with such problems as equipment heterogeneity, constraint dynamics and mission uncertainty. The traditional artificial intelligence method is difficult to solve the model uncertainty. The graph neural network technology is one of the connectionist artificial intelligence methods based on cognitive science, relational reinforcement learning. By building decision graph, by the decision graph vertex unmanned systems cluster state and constraint factors attribute of intelligent decision with the decision graph to represent the logic relations between various decision-making factors attribute, through reinforcement learning method to train the whole decision graph vertices attributes, attribute and decision graph of global properties, thereby graph by extracting decision attribute information to realize the underwater unmanned systems artificial intelligence decision-making. The purpose of this paper is to explore the graph neural network technology application in underwater unmanned systems intelligent aided decision-making feasibility of exploratory on underwater unmanned intelligent aided decision-making system for the graph neural network modeling, building intelligent auxiliary decision-making reasoning algorithm pseudo-code, research based on the technique of graph neural network underwater unmanned system implementation of the intelligent decision technique.
2020,42(12): 63-66 收稿日期：2020-08-18
DOI：10.3404/j.issn.1672-7649.2020.12.012
分类号：TP3-05；TP301.6
基金项目：四川省国际合作基金资助项目（2019YFH0017）
作者简介：冯振宇(1990-),男,博士研究生,研究方向为人工智能强化学习、图神经网络、无人系统集群智能决策
参考文献：
[1] HAMRICK J B, ALLEN K R, BAPST V, et al. Relational inductive bias for physical construction in humans and machines[J]. arXiv: Learning, 2018.
[2] WANG T, LIAO R, BA J, et al. NerveNet: Learning structured policy with graph neural networks[C]. International Conference on Learning Representations, 2018.
[3] BATTAGLIA P W, HAMRICK J B, BAPST V, et al. Relational inductive biases, deep learning, and graph networks[J]. arXiv: Learning, 2018.
[4] ZAMBALDI V, RAPOSO D, SANTORO A, et al. Relational deep reinforcement learning.[J]. arXiv: Learning, 2018.
[5] SILVER D, HUBERT T, SCHRITTWIESER J, et al. A general reinforcement learning algorithm that masters chess, shogi, and go through self-play[J]. Science, 2018, 362(6419): 1140-1144
[6] MNIH V, KAVUKCUOGLU K, SILVER D, et al. Human-level control through deep reinforcement learning[J]. Nature, 2015, 518(7540): 529-533
[7] ZHANG Y, CHEN X, YANG Y, et al. Can graph neural networks help logic reasoning[J]. arXiv: Learning, 2019.
[8] VINYALS O, BABUSCHKIN I, CZARNECKI W M, et al. Grandmaster level in StarCraft II using multi-agent reinforcement learning[J]. Nature, 2019, 575(7782): 350-354
