Symbiotic Evolution of Neural Networks in Sequential Decision Tasks

顺序决策任务出现在许多实际领域中，包括控制、资源分配、路由和调度。本课题的目标是开发一种基于神经网络共生进化的序列决策新方法。在共生进化中，一群神经元通过遗传算法进化，相互合作并形成决策网络。作为任务的一部分，保持种群的多样性，即使在稀疏强化的困难任务中，系统也能有效地找到好的解决方案。在拟议的项目中，共生进化将从理论上和实验上进行分析，该算法将进一步开发并应用于复杂的现实世界领域，如局域网管理，并将开发一个实用的高级接口，使该方法能够快速应用于新领域。本研究的主要科学贡献是：(1)为序列决策提供了一种新的、强大的方法来自动提取和编码特定问题的知识；(2)对多样性和合作在遗传算法中的作用有了透彻的理解。通用应用接口的发展也将使许多实际领域受益，包括控制工程、军事科学和运营管理。本文的主要假设是：(1)神经网络的模式识别和泛化能力可用于实现有效和鲁棒的序列决策策略；(2)遗传算法即使在稀疏强化下也能发现强大的针对特定问题的决策策略；3)通过使种群多样性成为任务的重要组成部分，共生进化可以为更难的问题找到解决方案，并且比标准遗传算法更有效。

Sequential decision tasks appear in many real-world domains, including control, resource allocation, routing, and scheduling. The objective of this project is to develop a new approach to sequential decision making based on symbiotic evolution of neural networks. In symbiotic evolution, a population of neurons are evolved with genetic algorithms to cooperate and form decision-making networks. Diversity is maintained in the population as part of the task and the system can find good solutions efficiently even in difficult tasks with sparse reinforcement. In the proposed project, symbiotic evolution will be analyzed both theoretically and experimentally, the algorithm will be further developed and applied to complex real-world domains such as local-area network management, and a practical high-level interface will be developed that will allow rapid application of the method to new domains. The main scientific contributions of the research are expected to be: (1) a novel, powerful method for extracting and encoding problem-specific knowledge automatically for sequential decision making, and (2) a thorough understanding of the role of diversity and cooperation in genetic algorithms. The development of the general application interface should also benefit many practical fields, including control engineering, military science, and operations management. The main hypotheses to be tested are: (1) Pattern recognition and generalization capabilities of neural networks can be used to implement effective and robust sequential decision making strategies;(2) Genetic algorithms can discover powepful problem-specific decision-making strategies even under sparse reinforcement; 3) By making population diversity an essential part of the task, symbiotic evolution can develop solutions to harder problems and do it more efficiently than standard genetic algorithms.