基于Q学习的震后救援路径寻优算法研究

发布时间：2018-02-12 10:09

本文关键词： 防灾减灾应急救援最优路径 Q学习智能体可靠性　出处：《石家庄铁道大学》2015年硕士论文　论文类型：学位论文

【摘要】：地震危害巨大,强烈的地震会使道路两侧建筑物倒塌从而阻断道路,使整个路网遭到破坏。当地震发生后城市外的救援人员需要将大量救援物资及时送往灾区,而城市内的人们则需要尽快将受伤的群众和灾民送往安全地区。因此,能否以最快的速度建立救援通道对于挽救受伤群众生命,减少财产损失显得至关重要。但是,城市内的一些基础设施由于受到地震的破坏,使电力、通信等系统部分甚至完全失效,导致灾区与外界在短时间内无法进行信息交流,这使得救援工作变的更加困难。本文根据人工智能的相关理论结合城市震后路网的实际特点,把每个救援队看作单个智能体,建立了基于Q学习的城市震后救援路径寻优模型。模型中以城市震后路网作为救援队所处环境,将路网中每个节点作为救援队的状态,救援队从节点到相邻节点的转移作为一个动作,把路段可靠度作为救援队学习的回报值。救援队经过一定次数的学习后可以得到每个状态-动作对的折扣累积的路段可靠度,即Q值。根据Q值大小,救援队确定动作选取的最优策略,通过该策略救援队可以找出一条到达灾区的最优路径。本文以长春市朝阳区部分路网为例进行了实例计算,验证了模型的可行性。并且,对模型参数进行了敏感性分析,结果表明当学习率上升,其它参数不变时,智能体学习速度加快;当折扣率上升,其它参数不变时,智能体学习速度降低。同时,考虑到地震发生后往往会存在多个救援队同时开展救援,建立了基于多智能体通信协作条件下的震后救援路径寻优模型。模型中以震后道路的破坏信息作为共享信息,救援队通过彼此信息的交流加快学习速度。由于在大地震发生后的最初阶段,通信设施遭到破坏,救援队间的通信很少,甚至完全没有。经过一定时间,通信率会逐渐增加。该模型中救援队之间的通信率初始值为零,随着时间的推进逐步增加。为了验证模型的可行性,同样以长春市朝阳区部分路网为例进行了实例分析,结果表明基于多智能体通信协作条件下的路径寻优模型可以以更快的速度找到一条较优的救援路径,多智能体的学习速度要比单智能体的学习速度提高了近三分之一。
[Abstract]:The earthquake is so dangerous that a strong earthquake can collapse buildings on both sides of the road and block the road and damage the entire road network. When the earthquake strikes, rescuers outside the city need to send a lot of relief supplies to the disaster area in time. People in the city need to send injured people and victims to safety as soon as possible. Therefore, it is important to establish rescue channels as quickly as possible in order to save the lives of the injured people and reduce the loss of property. Some of the infrastructure in the city was damaged by the earthquake, which completely invalidated some of the power, communication and other systems, resulting in a lack of information exchange between the disaster areas and the outside world in a short period of time. This makes the rescue work more difficult. According to the theory of artificial intelligence and the actual characteristics of the urban post-earthquake network, each rescue team is regarded as a single agent. Based on Q learning, the model of urban post-earthquake rescue path optimization is established. In the model, the post-earthquake road network is used as the environment of the rescue team, and each node in the road network is taken as the state of the rescue team. The transfer of rescue teams from nodes to adjacent nodes is taken as an action, and the section reliability is regarded as the reward value of the rescue team's learning. After a certain number of times of study, the rescue team can get the cumulative section reliability of each state-action pair discount. That is, Q value. According to the magnitude of Q value, the rescue team determines the optimal strategy of action selection, through which the rescue team can find out an optimal path to the disaster area. This paper takes part of the road network in Chaoyang District, Changchun City, as an example to carry out the calculation. The feasibility of the model is verified. The sensitivity analysis of the model parameters shows that when the learning rate increases and the other parameters remain unchanged, the agent learning speed accelerates, and when the discount rate increases, the other parameters remain the same. At the same time, considering that there are often more than one rescue team carrying out rescue work at the same time after an earthquake occurs, A model of post-earthquake rescue path optimization based on multi-agent communication cooperation is established, in which the damage information of the post-earthquake road is taken as the shared information. Rescue teams speed up their learning by exchanging information with each other. As communications facilities were destroyed in the early stages of the earthquake, there was little or no communication between rescue teams. After a certain period of time, The communication rate will increase gradually. In this model, the initial communication rate between rescue teams is zero, and gradually increases with the advance of time. In order to verify the feasibility of the model, a case study of some road networks in Chaoyang District, Changchun City, is also carried out. The results show that the path optimization model based on multi-agent communication cooperation can find a better rescue path at a faster speed. The learning speed of multi-agent is 1/3 faster than that of single agent.
【学位授予单位】：石家庄铁道大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：P315.9;U116.2;TP18

【相似文献】