基于历史梯度平均方差缩减的协同参数更新方法

随机梯度下降算法(SGD)随机使用一个样本估计梯度,造成较大的方差,使机器学习模型收敛减慢且训练不稳定。该文提出一种基于方差缩减的分布式SGD,命名为DisSAGD。该方法采用历史梯度平均方差缩减来更新机器学习模型中的参数,不需要完全梯度计算或额外存储,而是通过使用异步通信协议来共享跨节点的参数。为了解决全局参数分发存在的“更新滞后”问题,该文采用具有加速因子的学习速率和自适应采样策略：一方面当参数偏离最优值时,增大加速因子,加快收敛速度;另一方面,当一个工作节点比其他工作节点快时,为下一次迭代采样更多样本,使工作节点有更多时间来计算局部梯度。实验表明：DisSAGD显著减少了循环迭代的等待时间,加速了算法的收敛,其收敛速度比对照方法更快,在分布式集群中可以获得近似线性的加速。     

基于深度强化学习的机械臂控制快速训练方法

人工智能在机器人控制中得到广泛应用,机器人控制算法也逐渐从模型驱动转变为数据驱动。深度强化学习算法可在复杂环境中感知并决策,能够解决高维度和连续状态空间下的机械臂控制问题。然而,目前深度强化学习中数据驱动的训练过程非常依赖计算机GPU算力,且训练时间成本较大。提出基于深度强化学习的先简化模型（2D模型）再复杂模型（3D模型）的机械臂控制快速训练方法。采用深度确定性策略梯度算法代替机械臂传统控制算法中的逆运动学解算方法,直接通过数据驱动的训练过程控制机械臂末端到达目标位置,从而减小训练时间成本。同时,对于状态向量和奖励函数形式,使用不同的设置方式。将最终训练得到的算法模型在真实机械臂上进行实现和验证,结果表明,其控制效果达到了分拣物品的应用要求,相比于直接在3D模型中的训练,能够缩短近52%的平均训练时长。     

Continuous force field analysis for generalized gradient vector flow field

We propose a modification of the generalized gradient vector flow field techniques based on a continuous force field analysis. At every iteration the generalized gradient vector flow method obtains a new, improved vector field. However, the numerical procedure always employs the original image to calculate the gradients used in the source term. The basic idea developed in this paper is to use the resulting vector field to obtain an improved edge map and use it to calculate a new gradient based source term. The improved edge map is evaluated by new continuous force field analysis techniques inspired by a preceding discrete version. The approach leads to a better convergence and better segmentation accuracy as compared to several conventional gradient vector flow type methods.     

具有约束多目标优化的进化算法

针对带有约束多目标优化问题,提出一种多目标优化进化算法。在选择过程中,采用约束的Pareto支配和聚集距离定义适应值,根据适应值挑选出有代表性的个体。在变异过程中,沿着权重梯度方向搜索来寻找可行的Pareto最优解。最后,采用两个数值算例测草算法的性能,结果表明该算法能获得多目标约束优化问题的可行Pareto最优解并且具有较好的分散性。     

A local tangential lifting differential method for triangular meshes

In this note we present a local tangential lifting (LTL) algorithm to compute differential quantities for triangular meshes obtained from regular surfaces. First, we introduce a new notation of the local tangential polygon and lift functions and vector fields on a triangular mesh to the local tangential polygon. Then we use the centroid weights proposed by Chen and Wu [4] to define the discrete gradient of a function on a triangular mesh. We also use our new method to define the discrete Laplacian operator acting on functions on triangular meshes. Higher order differential operators can also be computed successively. Our approach is conceptually simple and easy to compute. Indeed, our LTL method also provides a unified algorithm to estimate the shape operator and curvatures of a triangular mesh and derivatives of functions and vector fields. We also compare three different methods : our method, the least square method and Akima’s method to compute the gradients of functions.     

Shape optimization for tumor location

In non-invasive thermal diagnostics, accurate correlations between the thermal image at skin surface and interior human physiology are desired. In this work, an estimation methodology to determine unknown geometrical parameters of an embedded tumor is proposed. We define a functional that represents the mismatch between a measured experimental temperature profile, which may be obtained by infrared thermography on the skin surface, and the solution of an appropriate boundary problem. This functional is related to the geometrical parameters through the solution of the boundary problem, in such a way that finding the minimum of this functional form also means finding the unknown geometrical parameters of the embedded tumor. Sensitivity analysis techniques coupled with the adjoint method were considered to compute the shape derivative of the functional. Then, a nonmonotone spectral projected gradient method was implemented to solve the optimization problem of finding the optimal geometric parameters.     

A near-optimal database allocation for reducing the average waiting time in the grid computing environment

In a grid computing environment, a great many users may access the same database simultaneously. To reduce the average waiting time for all users, a grid designer usually replicates the frequently accessed database among nodes based on the load balance heuristic. On the other hand, users may raise identical queries regarding an issue of interest, e.g., stock information, on a database and each of the queries will be directed to any node having a replica of that database. That is, the same answer will be determined by multiple nodes. Consequently, there exist two shortcomings of poor data sharing and duplicate calculations if the database is not replicated and allocated adequately. In this paper, we aim to minimize average waiting time and try to overcome the two shortcomings by performing database allocation over multiple nodes without any replication. The main idea behind the proposed method is to map the original problem to the Euclidean space Rⁿ and to solve the mapped problem in Rⁿ by a gradient-based optimization technique. The theoretical analyses ensure that the proposed method can converge linearly and achieve near-optimal results.     

Parameter estimation of Poisson mixture with automated model selection through BYY harmony learning

Finite mixture is widely used in the fields of information processing and data analysis. However, its model selection, i.e., the selection of components in the mixture for a given sample data set, has been still a rather difficult task. Recently, the Bayesian Ying-Yang (BYY) harmony learning has provided a new approach to the Gaussian mixture modeling with a favorite feature that model selection can be made automatically during parameter learning. In this paper, based on the same BYY harmony learning framework for finite mixture, we propose an adaptive gradient BYY learning algorithm for Poisson mixture with automated model selection. It is demonstrated well by the simulation experiments that this adaptive gradient BYY learning algorithm can automatically determine the number of actual Poisson components for a sample data set, with a good estimation of the parameters in the original or true mixture where the components are separated in a certain degree. Moreover, the adaptive gradient BYY learning algorithm is successfully applied to texture classification.     

Identification using ANFIS with intelligent hybrid stable learning algorithm approaches

This paper suggests novel hybrid learning algorithm with stable learning laws for adaptive network based fuzzy inference system (ANFIS) as a system identifier and studies the stability of this algorithm. The new hybrid learning algorithm is based on particle swarm optimization (PSO) for training the antecedent part and gradient descent (GD) for training the conclusion part. Lyapunov stability theory is used to study the stability of the proposed algorithm. This paper, studies the stability of PSO as an optimizer in training the identifier, for the first time. Stable learning algorithms for the antecedent and consequent parts of fuzzy rules are proposed. Some constraints are obtained and simulation results are given to validate the results. It is shown that instability will not occur for the leaning rate and PSO factors in the presence of constraints. The learning rate can be calculated on-line and will provide an adaptive learning rate for the ANFIS structure. This new learning scheme employs adaptive learning rate that is determined by input–output data.     

基于感兴趣区梯度方向直方图的行人检测

针对以梯度方向直方图作为人体特征的行人检测存在向量维数较大、检测时间较长的问题,提出基于感兴趣区梯度方向直方图的行人检测方法,分别在头部及四肢等重点区域计算梯度方向直方图,有效地减少了向量维数。实验结果表明,该方法在检测率基本不变的情况下提高了检测速度。