行动器评判器(Actor Critic,简称AC)算法是强化学习连续动作领域的一类重要算法,其采用独立的结构表示策略,但更新策略时需要大量样本导致样本效率不高.为了解决该问题,提出了基于模型学习和经验回放加速的正则化自然AC算法(Regularized...行动器评判器(Actor Critic,简称AC)算法是强化学习连续动作领域的一类重要算法,其采用独立的结构表示策略,但更新策略时需要大量样本导致样本效率不高.为了解决该问题,提出了基于模型学习和经验回放加速的正则化自然AC算法(Regularized Natural AC with Model Learning and Experience Replay,简称RNAC-ML-ER).RNAC-ML-ER将Agent与环境在线交互产生的样本用于学习系统动态性对应的线性模型和填充经验回放存储器.将线性模型产生的模拟样本和经验回放存储器中存储的样本作为在线样本的补充,实现值函数、优势函数和策略的更新.为了提高更新的效率,在每个时间步,仅当模型的预测误差未超过阈值时才利用该模型进行规划,同时根据TD-error从大到小的顺序对经验回放存储器中的样本进行回放.为了降低策略梯度估计的方差,引入优势函数参数向量对优势函数进行线性近似,在优势函数的目标函数中加入2-范数进行正则化,并通过优势函数参数向量来对策略梯度更新,以促进优势函数和策略的收敛.在指定的两个假设成立的条件下,通过理论分析证明了所提算法RNAC-ML-ER的收敛性.在4个强化学习的经典问题即平衡杆、小车上山、倒立摆和体操机器人中对RNACML-ER算法进行实验,结果表明所提算法能在大幅提高样本效率和学习速率的同时保持较高的稳定性.展开更多
Proper fixture design is crucial to obtain the better product quality according to the design specification during the workpiece fabrication. Locator layout planning is one of the most important tasks in the fixture ...Proper fixture design is crucial to obtain the better product quality according to the design specification during the workpiece fabrication. Locator layout planning is one of the most important tasks in the fixture design process. However, the design of a fixture relies heavily on the designerts expertise and experience up to now. Therefore, a new approach to loeator layout determination for workpieces with arbitrary complex surfaces is pro- posed for the first time. Firstly, based on the fuzzy judgment method, the proper locating reference and locator - numbers are determined with consideration of surface type, surface area and position tolerance. Secondly, the lo- cator positions are optimized by genetic algorithm（GA）. Finally, a typical example shows that the approach is su- perior to the experiential method and can improve positioning accuracy effectively.展开更多
Association rules and C4.5 rules can overcome the shortage of the traditional land evaluation methods and improve the intelligibility and efficiency of the land evaluation knowledge.In order to compare these two kinds...Association rules and C4.5 rules can overcome the shortage of the traditional land evaluation methods and improve the intelligibility and efficiency of the land evaluation knowledge.In order to compare these two kinds of classification rules in the application,two fuzzy classifiers were established by combining with fuzzy decision algorithm especially based on Second General Soil Survey of Guangdong Province.The results of experiments demonstrated that the fuzzy classifier based on association rules obtain a higher accuracy rate,but with more complex calculation process and more computational overhead;the fuzzy classifier based on C4.5 rules obtain a slightly lower accuracy,but with fast computation and simpler calculation.展开更多
[Objective]The aim was to overcome the shortage of being difficult to build land evaluation model when the impact factors had continuous value in the traditional land evaluation process,as well as to improve the intel...[Objective]The aim was to overcome the shortage of being difficult to build land evaluation model when the impact factors had continuous value in the traditional land evaluation process,as well as to improve the intelligibility of the land evaluation knowledge.[Method] The land evaluation method combining classification rule extracted by C4.5 algorithm with fuzzy decision was proposed in this study.[Result] The result of Second General Soil Survey of Guangdong Province had demonstrated that the method was convenient to extract classification rules,and by using only 100 rules,quantity correct rate 86.67% and area correct rate 84.80% of land evaluation could be obtained.[Conclusions] The use of C4.5 algorithm to obtain the rules,combined with fuzzy decision algorithm to build classifiers had got satisfactory results,which provided a practical algorithm for the land evaluation.展开更多
文摘行动器评判器(Actor Critic,简称AC)算法是强化学习连续动作领域的一类重要算法,其采用独立的结构表示策略,但更新策略时需要大量样本导致样本效率不高.为了解决该问题,提出了基于模型学习和经验回放加速的正则化自然AC算法(Regularized Natural AC with Model Learning and Experience Replay,简称RNAC-ML-ER).RNAC-ML-ER将Agent与环境在线交互产生的样本用于学习系统动态性对应的线性模型和填充经验回放存储器.将线性模型产生的模拟样本和经验回放存储器中存储的样本作为在线样本的补充,实现值函数、优势函数和策略的更新.为了提高更新的效率,在每个时间步,仅当模型的预测误差未超过阈值时才利用该模型进行规划,同时根据TD-error从大到小的顺序对经验回放存储器中的样本进行回放.为了降低策略梯度估计的方差,引入优势函数参数向量对优势函数进行线性近似,在优势函数的目标函数中加入2-范数进行正则化,并通过优势函数参数向量来对策略梯度更新,以促进优势函数和策略的收敛.在指定的两个假设成立的条件下,通过理论分析证明了所提算法RNAC-ML-ER的收敛性.在4个强化学习的经典问题即平衡杆、小车上山、倒立摆和体操机器人中对RNACML-ER算法进行实验,结果表明所提算法能在大幅提高样本效率和学习速率的同时保持较高的稳定性.
基金Supported by the Natural Science Foundation of Jiangxi Province(2009GZC0104)the Science and Technology Research Project of Jiangxi Provincial Department of Education(GJJ10521)~~
文摘Proper fixture design is crucial to obtain the better product quality according to the design specification during the workpiece fabrication. Locator layout planning is one of the most important tasks in the fixture design process. However, the design of a fixture relies heavily on the designerts expertise and experience up to now. Therefore, a new approach to loeator layout determination for workpieces with arbitrary complex surfaces is pro- posed for the first time. Firstly, based on the fuzzy judgment method, the proper locating reference and locator - numbers are determined with consideration of surface type, surface area and position tolerance. Secondly, the lo- cator positions are optimized by genetic algorithm（GA）. Finally, a typical example shows that the approach is su- perior to the experiential method and can improve positioning accuracy effectively.
基金Supported by Science and Technology Plan Project of Guangdong Province (2009B010900026,2009CD058,2009CD078,2009CD079,2009CD080)Special Funds for Support Program of Development of Modern Information Service Industry of Guangdong Province(06120840B0370124)Funded Fund Project of South China Agricultural University (2007K017)~~
文摘Association rules and C4.5 rules can overcome the shortage of the traditional land evaluation methods and improve the intelligibility and efficiency of the land evaluation knowledge.In order to compare these two kinds of classification rules in the application,two fuzzy classifiers were established by combining with fuzzy decision algorithm especially based on Second General Soil Survey of Guangdong Province.The results of experiments demonstrated that the fuzzy classifier based on association rules obtain a higher accuracy rate,but with more complex calculation process and more computational overhead;the fuzzy classifier based on C4.5 rules obtain a slightly lower accuracy,but with fast computation and simpler calculation.
基金Supported by Science and Technology Plan Project of Guangdong Province (2009B010900026,2009CD058,2009CD078,2009CD079,2009CD080)Special Funds for Support Program of Development of Modern Information Service Industry of Guangdong Province(06120840B0370124 )Fund Project of South China Agricultural University (2007K017)~~
文摘[Objective]The aim was to overcome the shortage of being difficult to build land evaluation model when the impact factors had continuous value in the traditional land evaluation process,as well as to improve the intelligibility of the land evaluation knowledge.[Method] The land evaluation method combining classification rule extracted by C4.5 algorithm with fuzzy decision was proposed in this study.[Result] The result of Second General Soil Survey of Guangdong Province had demonstrated that the method was convenient to extract classification rules,and by using only 100 rules,quantity correct rate 86.67% and area correct rate 84.80% of land evaluation could be obtained.[Conclusions] The use of C4.5 algorithm to obtain the rules,combined with fuzzy decision algorithm to build classifiers had got satisfactory results,which provided a practical algorithm for the land evaluation.
