A study of distributed intelligent systems for creation of future social welfare infrastructure

为了创造未来的社会福利基础设施，分布式智能的基础研究、模拟实验评估和应用的可行性研究已经得到了发展。本研究的现状是，智能算法可以开发出适合于福利和社会系统组成部分的可靠和低成本的设备。为了实现这些智能系统，本文提出了一种新的强化学习算法，采用多智能体设计方法，并实现了新的分布式智能应用。所提出的方法和技术使智能体能够根据环境在多个任务中获得自主自适应行为，并在传感器限制的情况下构建自适应状态空间以实现高效学习。具体来说，本研究带来了以下主要结果。(1)提出了一种低依赖传感器组态的有效状态空间构造方法。该方法包括用矩形对周围环境进行抽象和用自组织映射（SOM）进行分类对状态空间进行泛化。仿真实验表明，所构建的状态空间可以被具有其他传感器配置的智能体重用。(2)针对部分可观察马尔可夫决策过程环境下智能体的学习问题，提出了动作值为复数的q学习方法。预计该技术将使代理能够获得类似于复杂值神经网络的上下文依赖行为。两个实验的结果都表明，智能体有可能通过使用上下文来学习pomdp环境下的行为。(3)提出了一种基于分布式资源的基于多智能体的自主配电网恢复系统，该系统采用一种新的遗传算法和契约网络协议。所提出的系统使我们能够提高恢复性能，并减少计算资源和成本。仿真结果表明，在纯分布式环境下，所提方法达到了提高性能的目的。少

For creation of future social welfare infrastructure, fundamental studies of distributed intelligence, evaluations with simulation experiments, and feasibility studies of applications have been developed. The state of the art of this study is that the intelligent algorithm can develop dependable and low cost devices appropriate for the component of welfare and social systems. In order to realize these intelligent systems, this study proposes a new reinforcement learning algorithm, approaches with multi agent design, and implements of new distributed intelligent applications. The proposed methods and techniques enable agents to acquire autonomously adaptive behavior in several task according to environments, and to construct adaptive state space in spite of sensor limitations for efficient learning. Specifically the following main outcomes have been brought from this study.(1) A new method of constructing effective state space with low dependence on sensor configurations is conducted. T … More he method consists of abstracting surrounding environment by rectangles, and generalization of state space by classifying with self organization map(SOM). Simulation experiments show that the state space that has been constructed can be reused for the agent with another sensor configuration.(2) Q-learning with action values in complex numbers is proposed to overcome agent's learning under partially observable Markov decision processes environments. It is expected that the technique enables an agent to acquire context-dependent behaviors similar to complex-valued neural networks. The results of both experiments showed that there are possibilities for agents to learn behaviors under the POMDPs environments by using contexts.(3) A multi agent based autonomous power distribution network restoration system by using a new genetic algorithm and contract network protocol under the distributed resources is proposed. The proposed system enables us to improve the restoration performance, and to reduce computational resources and costs. The simulation results show the proposed method achieves to improve the performance under the pure distributed environment. Less