Complex Objective Functions Research Articles

Learning the feature distribution similarities for online time series anomaly detection

Identifying anomalies in multi-dimensional sequential data is crucial for ensuring optimal performance across various domains and in large-scale systems. Traditional contrastive methods utilize feature similarity between different features extracted from multidimensional raw inputs as an indicator of anomaly severity. However, the complex objective functions and meticulously designed modules of these methods often lead to efficiency issues and a lack of interpretability. Our study introduces a structural framework called SimDetector, which is a Local–Global Multi-Scale Similarity Contrast network. Specifically, the restructured and enhanced GRU module extracts more generalized local features, including long-term cyclical trends. The multi-scale sparse attention module efficiently extracts multi-scale global features with pattern information. Additionally, we modified the KL divergence to suit the characteristics of time series anomaly detection, proposing a symmetric absolute KL divergence that focuses more on overall distribution differences. The proposed method achieves results that surpass or approach the State-of-the-Art (SOTA) on multiple real-world datasets and synthetic datasets, while also significantly reducing Multiply-Accumulate Operations (MACs) and memory usage.

Comparative study of PI-controller and neurocontroller performances in optimal by settling time control problems

When developing control systems, an important issue arises of choosing an operator, which forms the control function. The standard approach is to use a PI- or PID-controller, a more advanced approach involves ANNs training for this purpose. A comparative analysis of the PI- and neurocontroller performances makes it possible to establish the disadvantages and advantages of each of the compared controllers, which is an important scientific and applied problem. The purpose of the work was to conduct a comparative analysis of the performance of the PI-controller and the neurocontroller based on a set of evaluation indicators for plants of the second and third orders. Such a comparison was carried out by using an approach to the synthesis of both controllers, which involved the minimization of a complex objective function. The latter is obtained as a result of reducing the problem of optimal control with constraints to the problem of unconstrained optimization. The analysis showed that according to the settling time indicator (optimization criterion), the neurocontroller has an advantage of 6.1...96.2% for the modelled plants. At the same time, according to other indicators of the control quality, the PI-controller has an advantage. In addition, the synthesis of a neurocontroller in terms of finding the minimum of the objective function is a more difficult problem. For its solution, a bigger number of iterations of the VCT-PSO optimization algorithm is required. It is rationally to set more than 1000 iterations and swarm population in the range 30…50 particles. A comparative analysis by the settling time of the neurocontroller and PI-controller, which is tuned according to engineering methods, showed significant reserves for improving this indicator. Thus, if the requirements for settling time minimization are quite strict, then it is advisable to use a neurocontroller. The obtained results will make it possible to develop recommendations for the rational choice of the control operator when solving practical problems of the control systems synthesis

Advancing environmental assessment methodologies: A case study on Pentle's framework through component analysis for complex objective functions

This paper explores the evolution and refinement of environmental assessment methods in response to contemporary challenges. While various methodologies exist, R. Pentle's proposed calculation method, represented by a multiple linear regression equation, stands as a significant contribution. However, this method requires refinement. To address this, the paper uses a case study approach within the State National Natural Park (SNNP) Katon-Karagay to investigate the impact of recreational load on tourist routes and the associated environmental load. Through component analysis, a complex objective function is developed within Pentle's framework, emphasizing the need for a standard-setting procedure and offering a fresh perspective on environmental functions. Rather than conventional rating scales, the study advocates for an innovative approach to delineating environmental scenarios. Importantly, this approach extends beyond environmental data analysis, offering applicability across diverse datasets, including economic and social inputs. The findings contribute to advancing environmental assessment methodologies, offering a versatile framework for understanding and managing the impacts of recreational and environmental loads in protected natural areas.

High‐throughput in silico workflow for optimization and characterization of multimodal chromatographic processes

AbstractWhile high‐throughput (HT) experimentation and mechanistic modeling have long been employed in chromatographic process development, it remains unclear how these techniques should be used in concert within development workflows. In this work, a process development workflow based on HT experiments and mechanistic modeling was constructed. The integration of HT and modeling approaches offers improved workflow efficiency and speed. This high‐throughput in silico (HT‐IS) workflow was employed to develop a Capto MMC polishing step for mAb aggregate removal. High‐throughput batch isotherm data was first generated over a range of mobile phase conditions and a suite of analytics were employed. Parameters for the extended steric mass action (SMA) isotherm were regressed for the multicomponent system. Model validation was performed using the extended SMA isotherm in concert with the general rate model of chromatography using the CADET modeling software. Here, step elution profiles were predicted for eight RoboColumn runs across a range of ionic strength, pH, and load density. Optimized processes were generated through minimization of a complex objective function based on key process metrics. Processes were evaluated at lab‐scale using two feedstocks, differing in composition. The results confirmed that both processes obtained high monomer yield (>85%) and removed of aggregate species. Column simulations were then carried out to determine sensitivity to a wide range of process inputs. Elution buffer pH was found to be the most critical process parameter, followed by resin ionic capacity. Overall, this study demonstrated the utility of the HT‐IS workflow for rapid process development and characterization.

Improvement of augmented free-water flash algorithm based on HV mixing rule and simulated annealing optimization algorithm

The calculation of three-phase equilibrium in CO2-hydrocarbon-water mixtures holds significant importance within numerical simulations, particularly in applications such as CO2-enhanced oil recovery and carbon dioxide sequestration. However, due to the non-ideality of CO2 and the polarity of water, three-phase equilibrium calculations often encounter convergence challenges. The augmented free-water flash algorithm [6], specifically focusing on CO2 dissolution in the aqueous phase, addresses the convergence challenges encountered by traditional three-phase flash algorithms. Despite its accuracy diminishes under conditions of high pressure and high CO2 concentrations, limiting its capability to accurately describe CO2 dissolution in oil-gas-water multiphase systems. The GE (excess Gibbs free energy) type mixing rule can be effectively integrated with equations of state, including PR and SRK, by utilizing activity models like NRTL and UNIFAC, which can be applied to the calculation of thermodynamic parameters for both nonpolar and polar systems and high-pressure complex systems. In this study, based on the augmented free-water assumption, we have developed a new augmented free-water three-phase flash algorithm for CO2/hydrocarbon/water mixtures. This algorithm integrates the PR equation of state with the NRTL activity model, employing the HV mixing rule for CO2-water interactions and the Van der Waals rule for other non-polar components. Furthermore, to enhance computational efficiency, the algorithm incorporates the simulated annealing global optimization algorithm, replacing Newton iteration to more effectively identify the global minima of the complex objective function. The example calculations show that the improved augmented free-water flash algorithm has higher accuracy and efficiency than the traditional augmented free-water flash algorithm. The augmented free-water three-phase flash algorithm proposed in this paper offers a precise model for phase equilibrium calculations essential for numerical simulations in CO2-enhanced oil recovery and carbon dioxide sequestration.

An efficient approach to parameter extraction of photovoltaic cell models using a new population-based algorithm

This article discusses the problem of accurate and efficient modeling of photovoltaic (PV) panels. It is a highly nonlinear problem. The following models were considered: a single diode model, a double diode model, a triple diode model, a four diode model, a module model (a poly-crystalline Photowatt-PWP201 module and a mono-crystalline STM6-40/36 module). The article presents a mathematical notation of these models, a detailed interpretation of their individual components, and a comparison of obtained results. To increase the effectiveness of modeling, a new population-based algorithm which can handle complex objective functions and a large number of decision variables was developed. This is important for the problem of identifying the parameters of PV cell models because each evaluation of the objective function requires calculating a set of points that determine the current–voltage characteristics. Moreover, in the considered problem a solution is searched with the use of the trial and error method. The proposed algorithm is called Micro Adaptive Fuzzy Cuckoo Search Optimization (μAFCSO). The μAFCSO algorithm uses several new mechanisms that were developed based on our experience with population-based algorithms. The use of these mechanisms has produced very good results in simulations. In the scope of simulation studies, the μAFCSO algorithm was used for parameter extraction in six PV cell models and was also applied to optimize fifteen typical test functions. The test functions were considered in order to demonstrate that our algorithm can be used to solve typical problems processed using population-based algorithms. The results obtained in this study were compared with the results obtained using well-established algorithms. The results obtained in this work are better or comparable to them.

Constraint Learning to Define Trust Regions in Optimization over Pre-Trained Predictive Models

There is a recent proliferation of research on the integration of machine learning and optimization. One expansive area within this research stream is optimization over pre-trained predictive models, which proposes the use of pre-trained predictive models as surrogates for uncertain or highly complex objective functions. In this setting, features of the predictive models become decision variables in the optimization problem. Despite a recent surge in publications in this area, only a few papers note the importance of incorporating trust-region considerations in this decision-making pipeline, that is, enforcing solutions to be similar to the data used to train the predictive models. Without such constraints, the evaluation of the predictive model at solutions obtained from optimization cannot be trusted and the practicality of the solutions may be unreasonable. In this paper, we provide an overview of the approaches appearing in the literature to construct a trust region and propose three alternative approaches. Our numerical evaluation highlights that trust-region constraints learned through our newly proposed approaches compare favorably with previously suggested approaches, both in terms of solution quality and computational time. History: Accepted by Pascal Van Hentenryck, Area Editor for Computational Modeling: Methods & Analysis. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2022.0312 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2022.0312 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ .

Super‐evolutionary mechanism and Nelder‐Mead simplex enhanced salp swarm algorithm for photovoltaic model parameter estimation

AbstractIn the pursuit of enhancing the efficiency of solar cells, accurate estimation of unspecified parameters in the solar photovoltaic (PV) cell model is imperative. An advanced salp swarm algorithm called the Super‐Evolutionary Nelder‐Mead Salp Swarm Algorithm (SENMSSA) is proposed to achieve this objective. The proposed SENMSSA addresses the limitations of SSA by incorporating a super‐evolutionary mechanism based on a Gaussian‐Cauchy mutation and a vertical and horizontal crossover mechanism. This mechanism enhances both the global optimization capabilities and the local search performance and convergence speed of the algorithm. It enables a secondary refinement of the global optimum, unlocking untapped potential in the solution space near the global optimum and elevating the algorithm's precision and exploitation capabilities to higher levels. The Nelder‐Mead simplex method is further introduced to enhance local search capabilities and convergence accuracy. The Nelder‐Mead simplex method is a versatile optimization algorithm that improves local search by iteratively adjusting a geometric shape (simplex) of points. It operates without needing derivatives, making it suitable for non‐smooth or complex objective functions. To assess the efficacy of SENMSSA, a comparative analysis is conducted against other available algorithms, namely SSA, IWOA, SCADE, LWOA, CBA, and RCBA, using the CEC2014 benchmark function set. Subsequently, the algorithm was employed to determine the unknown PV parameters under fixed conditions for three different diode models. Additionally, SENMSSA is utilized to estimate PV parameters for three commercially available PV models (ST40, SM55, KC200GT) under varying conditions. The experimental results indicate that the SENMSSA proposed in this study displays a remarkably competitive performance in all test cases compared to other algorithms. As such, we consider that the SENMSSA algorithm constitutes a reliable and efficient solution for the challenge of PV parameter estimation.

Optimal design of constant‐stress accelerated degradation tests based on the Wiener process with manufacturing batches heterogeneity and individual differences

AbstractThe use of Constant‐stress accelerated degradation tests (CSADT) modeling is an effective way to assess the reliability of products. However, when conducting reliability experiments, the data collected may come from various sources such as different manufacturing batches, equipment, and operators. This can result in reduced accuracy of the test evaluations. In this research paper, we propose optimal designs for the Wiener constant‐stress accelerated degradation model that take into account the heterogeneity of manufacturing batches and individual differences. Our goal is to minimize the variability of the mean time to failure (MTTF) while working within a specified budget, test time, and available test units. To achieve this, we utilize particle swarm optimization (PSO) to find the best solution based on a complex objective function. We also compare our model with others using degradation data from LEDs, showing that our model has a better goodness‐of‐fit. Finally, we present examples of optimal CSADT plans under different constraints and conduct sensitivity analysis.

Distributionally-robust chance-constrained optimization of selective maintenance under uncertain repair duration

The joint selective maintenance and repairperson assignment problem with uncertain duration of maintenance actions is studied in this paper. This problem arises in mission-oriented assets that perform sequences of missions interspersed with scheduled pauses to allow maintenance activities. The goal is to find the optimal subset of components to maintain and the maintenance level to be carried out by each repairperson that maximizes system reliability for the following mission. We first reformulate the complex objective function via single-variable concave functions before applying piecewise-linear approximation. This tight approximation enables large instances of the problem to be solved efficiently. We also consider the commonly encountered case of probabilistic maintenance duration with their distributions estimated from limited historical data samples. Distributionally-robust chance constraints (DRCCs) are employed to ensure that the maintenance tasks assigned to each repairperson can be completed within the break for a set of distributions contained in a Wasserstein-1 ball around the empirical distribution of the field data. A CVaR-based tractable approximation is applied to the DRCCs such that the problem can be reformulated into a mixed integer linear program readily handled by standard solvers. Results from the numerical experiments conducted on large benchmark instances demonstrated the computational efficiency of the proposed models and the added value of considering distributional ambiguity when dealing with uncertain maintenance duration.

An improved inversion method for shale elastic parameters measurement based on ultrasonic resonance spectroscopy

The accurate measurement of elastic parameters of shale cores has important guidance for the exploration and development of shale oil and gas. The traditional shale acoustic experiment requires multiple samples from different angles and its preparation process is complex. Resonant Ultrasound Spectroscopy has the advantages of repeatable measurement, high accuracy, and the ability to obtain all elastic parameters from a single small sample. However, the influence of clamping force and measuring angle needs comprehensive analysis to ensure that formant can be measured as many and accurately as possible. This research first conducts sensitivity analysis on the elastic parameters of different shale to guide the experiment. In addition, due to the more complex inversion objective function of the anisotropic shale compared to isotropic samples, the traditional Levenberg–Marquardt method has a strong dependence on the initial value and is prone to falling into local optimal solution. The L-BFGS (limited memory BFGS) method is introduced into the inversion of anisotropic elastic parameters of shale, which can significantly improve the inversion efficiency. Finally, the robustness and reliability of the inversion method are validated through theoretical and experimental measurement results. [Work supported by the National Natural Science Foundation of China (42004117 and 42127807).]

Inversion analysis of rock mass permeability coefficient of dam engineering based on particle swarm optimization and support vector machine: A case study

In order to efficiently and reasonably determine the permeability coefficient of engineering rock mass, the two key problems of complex seepage field simulation and objective function solution in permeability coefficient inversion are studied and improved. Firstly, a 3D refined seepage numerical model is established considering the main fault fracture zones and dam seepage control system. Secondly, combined with the support vector machine (SVM) algorithm, the complex mapping relationship between permeability coefficient of each rock layer and impervious structure and piezometer head is constructed. Finally, the particle swarm optimization (PSO) algorithm is introduced, and then the efficient optimization of permeability coefficient inversion of engineering rock mass and impervious structure is realized based on PSO-SVM. The case study of LJH arch dam shows that the inversion calculation framework based on PSO-SVM not only completes the rapid proxy calculation of the numerical model of complex seepage field, but also realizes the accurate optimization of the inversion objective function. The simulation results under the final inversion parameters reproduce the variation characteristics of the piezometric tube water level of each typical measuring point under the change of reservoir water level. In addition, with the correlation analysis between the permeability coefficient and the measured value of the corresponding piezometer, the parameter inversion method is endowed with new application value. This study is of great significance for clarifying the seepage characteristics of complex rock mass and improving the inversion efficiency of dam seepage monitoring data.

Feature selection with multi-class logistic regression

Feature selection can help to reduce data redundancy and improve algorithm performance in actual tasks. Most of the embedded feature selection models are constructed based on square loss and hinge loss. However, these models based on the square loss cannot directly evaluate the discriminability of the samples in the feature subspace, and these methods based on the hinge loss are difficult to solve due to their complex objective functions. To deal with these problems, a Feature Selection method with Multi-class Logistic Regression (FSMLR) is proposed in this paper. Firstly, we construct a linear function to measure the difference between the distance from samples to their regression hyperplane and the distance from these samples to regression hyperplanes of other classes, which could be used to strengthen the discriminant property of the embedded model. Then, we design a re-weighting matrix with a ℓ2,0-norm sparse condition as well as a discrete condition, which is used to select features in the subspace. Considering that it is difficult to solve the re-weighting matrix with the discrete and sparse conditions in an optimization problem, we relax these two conditions and present a feature selection model via a re-weighted multi-class logistic regression with the two relaxed constraints. Finally, we add the F-norm regularization in our model to avoid overfitting, and its unconstrained equivalent transformation with ℓ2,p-norm regularization is derived to explore the function of the re-weighting matrix. The gradient descent algorithm could be used to solve the FSMLR. Especially, when the regularization term in the equivalence problem is set to ℓ2,1-norm, the global optimal solution can be obtained. Extensive experiments on multiple public data sets prove that FSMLR outperforms other competitors.

A comparative study of metaheuristics algorithms based on their performance of complex benchmark problems

Metaheuristic approaches with extremely important improvements are very promising in the solution of intractable optimization problems. The objective of the present study is to test the capability of applications and compare the performance of the four selected algorithms from “classical” (simulated annealing (SA), genetic algorithm (GA), particle swarm optimization (PSO), and differential evolution (DE)) and “new generation” (firefly algorithm (FFA), krill herd (KH), grey wolf optimization (GWO), and symbiotic organism search (SOS)) each by solving selected benchmark problems that are used in the literature for algorithm testing purpose. The selected test problems had very complex objective functions and associated constraints with multiple local optima. Among all selected algorithms, the “new generation” SOS and KH algorithm successfully solved most of all the selected benchmark problems and achieved the best solution for most of them. Among four “classical” algorithms, DE, and PSO effectively attained the optimal solution which was very close to the best one. However, the “new generation” algorithm performed much better than the “classical” one. Therefore, no firm conclusion can be done about the universally best algorithm and their performance may be varied for different benchmark problems. However, in this study for the seven selected test problems, SOS and KH exhibited the most promising result and great potential with respect to execution time also. This study gives some insights to use SOS and KH as the best-performing algorithms to the novice user who can easily get lost in the plethora of large optimization algorithms.

UAV Trajectory and Energy Efficiency Optimization in RIS-Assisted Multi-User Air-to-Ground Communications Networks

An air-to-ground downlink communication network consisting of a reconfigurable intelligent surface (RIS) and unmanned aerial vehicle (UAV) is proposed. In conjunction with a resource allocation strategy, the system’s energy efficiency is improved. Specifically, the UAV equipped with a RIS starts from an initial location, and an energy-efficient unmanned aerial vehicle deployment (EEUD) algorithm is deployed to jointly optimize the UAV trajectory, RIS phase shifts, and BS transmit power, so as to obtain a quasi-optimal deployment location and hence improve the energy efficiency. First, the RIS phase shifts are optimized by using the block coordinate descent (BCD) algorithm to deal with the nonconvex inequality constraint, and then integrated with the Dinkelbach algorithm to address the resource allocation problem of the BS transmit power. Finally, for solving the UAV trajectory optimization problem, the complex objective function is transformed into a convex function, and the optimal UAV flight trajectory is obtained. Our simulation results show that the quasi-optimal deployment location obtained by the EEUD algorithm is superior to other deployment strategies in energy efficiency. Moreover, the instantaneous energy efficiency of the UAVs along the trajectory of searching the deployment location is better than other comparison trajectories. Furthermore, the RIS-assisted multi-user air-to-ground communication network can offer up to 145% improvement in energy efficiency over the traditional amplify-and-forward (AF) relay.

Multi-Objective Parametric Optimization Design for Mirrors Combined with Non-Dominated Sorting Genetic Algorithm

The process of intelligent multi-objective parametric optimization design for mirrors is discussed in detail in this paper, with the error of the mirror surface shape and the total mass being examined as the optimization objectives. The establishment of complex objective functions for solving the optimization problem of the mirror surface shape error was realized, and manual modification of the model was avoided. Moreover, combining this with a non-dominated sorting genetic algorithm (NSGA) helped the Pareto front move towards an ideal optimal set of solutions. To verify the effectiveness of the proposed method, an aluminum alloy mirror with an aperture of 140 mm was taken as an example. The Pareto optimal solution set of the mass and surface shape error under 1 g gravity was obtained for finding the required solution and satisfying the optimization goal. In addition, this method is applicable to other complex structural design problems.

Food inspector scheduling with outcome and daily-schedule effects

Food-safety inspection is regularly executed by the government for quality assessment. Evidence from recent research demonstrates that inspection accuracy and consistency are affected by inspection biases that result from an operational decision: inspector scheduling. More precisely, an inspector's stringency in an inspection is affected by the inspection results at the previous-inspected establishment (outcome effects) and when this inspection occurs within a workday (daily-schedule effects). To our best knowledge, the impact of these effects on scheduling decisions has not been studied in the scheduling literature. In this paper, we study a novel food inspector scheduling problem with these effects, where the inspector should scrutinise establishments with different locations. The problem is viewed as a single-machine scheduling problem with a complex objective function including (i) inspection accuracy, (ii) inspection consistency and (iii) workload of the inspector. To facilitate quantitative analyses of these effects, we model them by sequence-dependent functions and formulate a mixed integer linear programming model. To overcome the computational difficulty in large-scale problems, an efficient Tabu Search algorithm is developed. Experiment results on 135 randomly generated instances with up to 50 establishments and 10 workdays validate the efficiency of the solution method. Besides, managerial insights are drawn.

LEAP: Scaling Numerical Optimization Based Synthesis Using an Incremental Approach

While showing great promise, circuit synthesis techniques that combine numerical optimization with search over circuit structures face scalability challenges due to a large number of parameters, exponential search spaces, and complex objective functions. The LEAP algorithm improves scaling across these dimensions using iterative circuit synthesis, incremental reoptimization, dimensionality reduction, and improved numerical optimization. LEAP draws on the design of the optimal synthesis algorithm QSearch by extending it with an incremental approach to determine constant prefix solutions for a circuit. By narrowing the search space, LEAP improves scalability from four to six qubit circuits. LEAP was evaluated with known quantum circuits such as QFT and physical simulation circuits like the VQE, TFIM, and QITE. LEAP can compile four qubit unitaries up to 59× faster than QSearch and five and six qubit unitaries with up to 1.2× fewer CNOTs compared to the QFAST package. LEAP can reduce the CNOT count by up to 36×, or 7× on average, compared to the CQC Tket compiler. Despite its heuristics, LEAP has generated optimal circuits for many test cases with a priori known solutions. The techniques introduced by LEAP are applicable to other numerical optimization based synthesis approaches.

Extreme gradient boosting-inspired process optimization algorithm for manufacturing engineering applications

Design and process optimization are key aspects of manufacturing engineering. This contribution details a machine learning (ML) methodology capable of learning from simulation results, experimental data, or sensor signals, and able to predict and optimize specific user-defined process and design parameters. The prediction-optimization algorithm is based on an enhanced Extreme Gradient Boosting (XGB) algorithm, highly responsive and accurate even for large datasets and complex variable interactions. Once the XGB function is trained, the optimization is carried out by a metaheuristic search algorithm defined according to the Differential Evolution (DE) architecture, allowing for the definition of the combination of independent variables that grant the minimization, or maximization, of the user’s defined objective function. The prediction and optimization capabilities have been assessed by applying the XGB-DE algorithm to a previously published authors' dataset and experimental results relevant for the radial-axial ring rolling (RARR) process, showing a prediction accuracy and optimization capability equal to 2.33% and 27.4%, respectively, with respect to authors’ previous finite element and experimental results. The XGB-DE methodology showed a remarkable capability in catching the trend and global minimum of a multi-variable and complex objective function, such as the one involved in complex thermo-mechanical forming processes.

Optimization of a Regional Marine Environment Mobile Observation Network Based on Deep Reinforcement Learning

The observation path planning of an ocean mobile observation network is an important part of the ocean mobile observation system. With the aim of developing a traditional algorithm to solve the observation path of the mobile observation network, a complex objective function needs to be constructed, and an improved deep reinforcement learning algorithm is proposed. The improved deep reinforcement learning algorithm does not need to establish the objective function. The agent samples the marine environment information by exploring and receiving feedback from the environment. Focusing on the real-time dynamic variability of the marine environment, our experiment shows that adding bidirectional recurrency to the Deep Q-network allows the Q-network to better estimate the underlying system state. Compared with the results of existing algorithms, the improved deep reinforcement learning algorithm can effectively improve the sampling efficiency of the observation platform. To improve the prediction accuracy of the marine environment numerical prediction system, we conduct sampling path experiments on a single platform, double platform, and five platforms. The experimental results show that increasing the number of observation platforms can effectively improve the prediction accuracy of the numerical prediction system, but when the number of observation platforms exceeds 2, increasing the number of observation platforms will not improve the prediction accuracy, and there is a certain degree of decline. In addition, in the multi-platform experiment, the improved deep reinforcement learning algorithm is compared with the unimproved algorithm, and the results show that the proposed algorithm is better than the existing algorithm.