【作者】 李颖；

【导师】 吴建设；

【作者基本信息】 西安电子科技大学 ， 电路与系统， 2021， 硕士

【摘要】 随着智能化时代的来临,无人系统在农业、工业、航空、航海、军事等领域应用广泛。如今实际环境复杂多变,不同的任务需要用不同的算法实现且每个算法各有优缺点,单一的算法已不能满足实际需求。根据任务规划实现算法协同不仅能够利用算法其自身优势,而且可实现算法间的优势互补,提高无人系统的效率和智能化水平。本文主要研究无人系统中如何根据任务规划调用合适的算法以提高协同性和任务效率,主要工作如下:1.针对无人系统任务规划算法与特定任务需求难以匹配的问题,提出了基于特征表示的算法协同框架。在该框架下,建立算法特征与实验物理属性之间的映射关系,将算法和任务需求用相应的特征向量表示,构建算法表征库以实现算法协同调度。在编队分群和路径规划中验证了算法协同框架的有效性。2.Q-Learning算法在“探索未知环境”和“利用已知信息”之间无法平衡,导致难以得到最佳路径。据此,提出了基于路径规划特征表示的改进Q-Learning算法,算法中将?-贪心策略和玻尔兹曼策略相结合,对路径中的危险节点和转弯进行适当的惩罚。实验表明,该算法可得到转弯更少、长度更短且安全系数更高的路径,综合考虑有效Q值数目和收敛速度时具有显著优势。3.针对遗传算法随机初始化路径过长,难以得到最佳路径的问题,提出了基于路径规划特征表示的改进遗传算法。算法在构建路径时缩小了节点选择区域以避免路径过长,将精英选择策略加入到选择算子中以保留每代最佳路径,在适应度函数中对路径中的危险节点和转弯进行惩罚以优化路径。实验对比表明,改进遗传算法的运行时间有所减少且规划的路径质量更好。综上,将基于路径特征表示的改进Q-Learning算法和改进遗传算法应用到算法协同框架中,实验表明特定任务基于算法协同框架得到的路径质量有所提高。更多还原

【Abstract】 With the advent of the intelligent age,unmanned systems are widely used in agriculture,industry,aviation,navigation,military and other fields.Nowadays,the actual environment is complex and changeable.Different tasks need to be implemented with different algorithms and each algorithm has its own advantages and disadvantages.A single algorithm can no longer meet the actual needs.According to the task planning,implementing algorithm collaboration can not only take advantage of the algorithms’ own advantages,but also achieve the complementary advantages between algorithms,and improve the efficiency and intelligence of unmanned systems.This thesis mainly studies how to select an appropriate algorithm for task planning to improve collaboration and task efficiency in unmanned systems.The main work of this thesis is as follows:1.Aiming at the problem that the task planning algorithm of unmanned systems is difficult to match the requirements of specific tasks,an algorithm collaboration framework based on feature representation is proposed.Under this framework,the mapping relationship between algorithm features and experimental physical attributes is established,the algorithm and task requirements are represented by corresponding feature vectors,and the algorithm characterization library is constructed to accomplish the collaborative scheduling of algorithms.The effectiveness of the algorithm collaboration framework is verified in the formation grouping and path planning.2.It is difficult to obtain the best path for Q-Learning algorithm,because it cannot balance between "exploring the unknown environment" and "using known information".An improved Q-Learning algorithm based on the feature representation of path planning is proposed.The algorithm combines the ?-greedy strategy and Boltzmann strategy,and appropriately penalizes dangerous nodes and turns in the path.Experiments show that the algorithm can obtain paths with fewer turns,shorter lengths and higher safety factors,and it has significant advantages when considering the number of effective Q values and convergence speed.3.Aiming at the problem that randomly initialized paths are too long to get the best path in genetic algorithm,an improved genetic algorithm based on the feature representation of path planning is proposed.The algorithm reduces the node selection area when constructing the path to avoid the path from being too long.It adds an elite selection strategy to the selection operator to retain the best path of each generation,and penalizes dangerous nodes and turns in the path in the fitness function to optimize the path.Experimental comparison shows that the running time of the improved genetic algorithm is reduced and the quality of the planned path is better.In summary,the improved Q-Learning algorithm and improved genetic algorithm based on path feature representation,are applied to the algorithm collaboration framework.Experiments show that the quality of the path obtained for the specific task based on the algorithm collaboration framework has been improved.更多还原