Fast generation method of ship conceptual design scheme solution set based on multi-objective optimization

Abstract

This invention discloses a rapid method for generating a solution set of ship conceptual design schemes based on multi-objective optimization, comprising the following steps: 1) obtaining the parameters of the conceptual design scheme and the value range of each parameter according to the requirements of the ship conceptual design scheme; 2) establishing a differential evolution optimization model and using differential evolution to optimize the ship conceptual design scheme; 3) using a reinforcement learning model to optimize the parameters of the differential evolution optimization model; 4) using the optimized parameters to obtain the optimized differential evolution optimization model; 5) using the optimized model to perform overall performance optimization on the principal dimensions and hull lines data in the ship conceptual design scheme, and outputting the final solution set of the conceptual scheme. This invention uses an optimization method combining differential evolution and reinforcement learning, which can reduce the dimensionality of the states encountered during the reinforcement learning process, thereby making the rewards in the entire interaction process less sparse and greatly improving the overall performance of the differential evolution algorithm.

Description

Technical Field

[0001] This invention relates to ship-aided design technology, and more particularly to a method for rapidly generating a solution set of ship conceptual design schemes based on multi-objective optimization. Background Technology

[0002] Conceptual design of ships is the most important component of overall surface ship design. Current conceptual design prioritizes optimizing individual performance indicators, using other performance characteristics as constraints, and achieves a globally feasible solution through multiple rounds of convergence iteration. However, this design approach neglects the interactions between various performance characteristics and fails to effectively utilize the coupling effects between multiple disciplines. The resulting solution is merely one of many feasible solutions or the optimal solution for a single performance characteristic, rather than the overall optimal solution sought in conceptual design.

[0003] Traditional conceptual design methods struggle to fully account for complex coupling relationships and face computational difficulties. Furthermore, traditional design relies heavily on historical data and expert experience, resulting in subjective limitations and low efficiency. Therefore, there is an urgent need for a multi-objective optimization and multi-attribute decision-making intelligent method that pursues the overall optimal solution during the ship conceptual design phase. This method would effectively address the limitations, computational difficulties, and inefficiencies of traditional conceptual designs, providing a conceptual design scheme that satisfies overall optimality for the ship's overall design. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a rapid generation method for the solution set of ship conceptual design schemes based on multi-objective optimization, which addresses the deficiencies in the existing technology.

[0005] The technical solution adopted by this invention to solve its technical problem is: a rapid generation method for ship conceptual design scheme solution set based on multi-objective optimization, comprising the following steps:

[0006] 1) Obtain the parameters of the conceptual design scheme and the range of values ​​for each parameter according to the requirements of the ship conceptual design scheme;

[0007] The parameters include principal dimension parameters and profile parameters;

[0008] 2) Establish a differential evolution optimization model and use differential evolution to optimize the ship's conceptual design scheme;

[0009] The input to the differential evolution optimization model is the principal scale parameter and the profile data. Based on the hyperparameters represented by the next state, the model optimizes the current input conceptual design scheme and finally obtains the optimized conceptual design scheme.

[0010] 3) Optimize the parameters of the differential evolution optimization model;

[0011] The hyperparameters of differential evolution include: population size NP, scaling factor F, and crossover probability CR;

[0012] 4) Optimize the differential evolution algorithm model using the optimized parameters; after obtaining a set of hyperparameter combinations, calculate the overall performance evaluation value of the conceptual design output by the differential evolution optimization model under this set of hyperparameters;

[0013] 5) Use the optimized model to perform overall performance optimization on the main dimensions and hull lines data in the conceptual design scheme of the ship, and output the final conceptual scheme solution set.

[0014] According to the above scheme, step 3) of optimizing the parameters of the differential evolution optimization model includes the following steps:

[0015] 3.1) Initialize hyperparameters;

[0016] 3.2) Determine the range of each of the three hyperparameters of differential evolution as the search space, and evaluate the reward value given by the environment for the action when choosing an action in the current state to reach the next state in the search space to update the three hyperparameters.

[0017] 3.3) Optimize the current input conceptual design scheme based on the hyperparameters represented by the next state, and then input the next state into the optimization network model for the next round of iteration to train the optimization network model.

[0018] According to the above scheme, the reinforcement learning network structure in step 3) includes an input layer, an output layer, and a 3-layer LSTM network structure, wherein each layer of the Long Short-Term Memory (LSTM) network has 35 neurons; the output layer is composed of a softmax function; and there is a fully connected layer between the input layer, the output layer, and the 3-layer LSTM network structure.

[0019] The beneficial effects of this invention are:

[0020] 1. The method of this invention uses an optimization method that combines differential evolution and reinforcement learning, which can reduce the dimensionality of the states faced in the reinforcement learning process, thereby making the rewards in the entire interaction process less sparse and greatly improving the overall performance of the differential evolution algorithm. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0022] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0023] Figure 2 This is a flowchart of the reinforcement learning network training phase according to an embodiment of the present invention;

[0024] Figure 3 This is a diagram of the reinforcement learning network structure according to an embodiment of the present invention;

[0025] Figure 4 This is an optimization flowchart of the differential evolution optimization model according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0027] like Figure 1 As shown, a rapid method for generating solution sets of ship conceptual design schemes based on multi-objective optimization includes the following steps:

[0028] 1) Obtain the parameters of the conceptual design scheme and the range of values ​​for each parameter according to the requirements of the ship conceptual design scheme;

[0029] The parameters include principal dimension parameters and profile parameters;

[0030] There are five main dimensional parameters: length, beam, depth, draft, and block coefficient. There are twenty-three hull line parameters: inlet length, parallel midbody length, outlet length, longitudinal position of buoyancy center, rhombus coefficient, half inlet angle, waterline coefficient, longitudinal stern slope, bow inclination angle, bulbous bow length, bulbous bow width, height above bulbous bow baseline, height below bulbous bow baseline, stern inclination angle, stern tangent position, side inclination angle, side inclination angle, side inclination angle, side inclination angle, side inclination angle, side inclination angle, side inclination angle, side inclination angle, side inclination angle, side inclination angle, side inclination angle.

[0031] 2) Establish a differential evolution algorithm model and use differential evolution to optimize the ship conceptual design scheme;

[0032] The input to the differential evolution algorithm model is the principal scale parameter and the profile data. It optimizes the current input conceptual design scheme based on the hyperparameters represented by the next state, and finally obtains the optimized conceptual design scheme, which is the adjusted comprehensive performance optimized conceptual design parameter.

[0033] 3) Optimize the parameters of the differential evolution algorithm model;

[0034] The hyperparameters of differential evolution include: population size NP, scaling factor F, and crossover probability CR;

[0035] The hyperparameters of the differential evolution algorithm are controlled by reinforcement learning. The input of reinforcement learning is the hyperparameters of differential evolution, which include population size NP, scaling factor F, and crossover probability CR. The output is the optimal hyperparameters that enable differential evolution to obtain the conceptual design sample with the best overall performance.

[0036] During parameter optimization, the initial solution set of the ship concept design is first added based on the population size NP. Then, the individual samples in the solution set are optimized and selected according to the scaling factor F and the crossover probability CR, and simultaneously fed back to the reinforcement learning network for training. The overall algorithm is implemented in three stages: preparation, training, and inference, as detailed below:

[0037] 3.1) Preparation stage

[0038] The preparation phase includes two main parts: agent and environment design for reinforcement learning, and the design of methods for optimizing agent-environment interaction. First, the agent and environment are designed. In reinforcement learning, the environment design treats the three hyperparameters of differential evolution—population size NP, scaling factor F, and crossover probability CR—as coordinates in a three-dimensional space, using the ranges of each hyperparameter as the search space. All interactions between the agent and the environment occur within this search space. The agent's state in the environment is represented by these three hyperparameters. The initial state is the default value of the hyperparameters set by differential evolution. The initial value of population size NP is 5 × the ship parameter dimension (D, D = 28), with a range of values... The initial value of the scaling factor is 0.5, and its range is [missing value]. The initial value of the crossover probability CR is 0.1, and its range is [missing value]. State S can be represented by the following formula:

[0039] S = [NP F CR]

[0040] The agent's actions are determined by different hyperparameter settings. For the population size NP, its value needs to be a positive integer, representing the total number of ship concept design samples in the initial situation. Therefore, the action corresponding to this hyperparameter is ±1, as shown in formula A1. The scaling factor F and the crossover probability CR have the same dimensions, so their actions can be the same, with an action of ±0.01, as shown in formulas A2 and A3.

[0041]

[0042]

[0043]

[0044] Therefore, choosing one of the six actions based on the current state will lead to the next state. The state transition equation can be expressed as:

[0045] S i+1 =S i +A1[m]+A2[n]+A3[k],m,n,k∈{0,1}

[0046] Among them, S i+1 The next state is S. i A1[m] is the current state, and A2[n] is an action of the population size NP chosen in the current state, which is the data in one row. Similarly, A2[n] is an action of the scaling factor F, and A3[k] is an action of the crossover probability CR.

[0047] The interaction design between the agent and the environment involves the reward value R given by the environment to the agent when it chooses an action in the current state and reaches the next state. In other words, the environment evaluates the appropriateness of choosing the action in the current state based on the next state, and the evaluation result is reflected in the reward. To ensure that the conceptual design scheme obtained by differential evolution based on this set of hyperparameters is optimal overall, the reward value of this method is set as the hyperparameter represented by the next state. The difference between the comprehensive performance evaluation score of the conceptual design obtained by differential evolution and the performance evaluation score of the conceptual design obtained by the hyperparameter represented by the initial state is shown in the following formula:

[0048] R = Func(S) i+1 -Func(S0)

[0049] Here, Func(·) represents the performance evaluation function of the conceptual design scheme obtained through differential evolution. The performance evaluation score of the conceptual design scheme is calculated based on the ship parameters obtained from the conceptual design, comprehensively determining five performance indicators for the ship under those parameters: total resistance, Beaufort grade, turning diameter, displacement, and initial metacentric height. The final comprehensive performance evaluation score is obtained by normalizing these five performance indicators and then taking a weighted average. The iteration terminates when the absolute value of the reward exceeds a fixed threshold, currently set at 0.15.

[0050] 3.2) Training Phase

[0051] The overall process of the training phase is as follows: Figure 2As shown, the process begins by initializing the hyperparameters of differential evolution as the initial state for reinforcement learning. Then, the agent's neural network determines the action in the current state, and the next state is derived from the current action. The inputs to differential evolution are the principal scale parameters and profile data. Based on the hyperparameters representing the next state, the current input conceptual design is optimized, resulting in an optimized conceptual design. This design is then comprehensively evaluated to obtain a comprehensive score. The reward for choosing this action in the current state is obtained by subtracting the comprehensive score of the input conceptual design from the current comprehensive score. The next state is then input into the agent network for the next iteration.

[0052] The agent network in reinforcement learning of this invention is as follows: Figure 3 As shown, its core consists of a 3-layer Long Short-Term Memory (LSTM) network, with 35 neurons in each layer; the output layer is composed of a softmax function; and there is a fully connected layer between the input layer, the output layer, and the 3-layer LSTM network structure. The structure and parameters of the 3-layer LSTM network structure in the agent network are shared at any time.

[0053] In this invention, the agent model modifies only one hyperparameter at each time step, thus producing different outputs at different times. For example, at time 1, we optimize the population size NP based on the network. Given the current state, the agent network can output the probability values ​​of two actions in action set A1 to decide which action to choose. The current state plus the chosen action serves as the input at time 2, which is then used to decide the scaling factor, and so on. After all hyperparameters in the state have been updated, they are used as the agent's output to calculate the current reward. Simultaneously, all hull line parameters of the ship's conceptual design are updated.

[0054] Reinforcement learning models are trained using policy gradients. The optimization objective of the algorithm is to maximize the expected total reward value, which can be calculated using the following formula: Let τ be the set of all states and actions in a given time interval from 0 to 3 (called a trajectory), then:

[0055]

[0056] In the formula, r(τ) calculates the reward for trajectory τ, p θ (τ) represents the probability value that the agent network generates the trajectory through action, θ represents the parameters of the neural network in reinforcement learning, and the overall meaning of the formula is that the total reward value is the expected value of the reward for each state in a trajectory.

[0057] Since the optimization objective is to find the neural network parameters θ that maximize the expected overall reward, the gradient descent algorithm can be used to find the local optimum by calculating the gradient of the objective function and updating the parameters θ.

[0058]

[0059] According to the above formula, it can be seen that For function Expectation, π θ (a t |s k This indicates that action a is selected based on the agent network in the current state. t The probability of [the probability]. This intellectual achievement, under a fixed parameter θ, uses the mean of m samples as an unbiased estimate for gradient updates; m is typically taken as 64.

[0060]

[0061] Where r(τ) k ) represents the reward value obtained from the hyperparameter combination of the kth sampling; b is the baseline value, which is the exponential moving average of the rewards of the sampled algorithm model.

[0062] After obtaining a set of hyperparameter combinations, it is necessary to calculate the overall performance evaluation of the conceptual design under this set of hyperparameters using differential evolution optimization. The optimization process of the conceptual design scheme by differential evolution is as follows: Figure 4 As shown.

[0063] This invention uses a population-based objective optimization algorithm (DE) to obtain the required parameter model through heuristic exploration in an existing ship type sample space. Finally, the optimal individual, which is the sample parameter with the highest comprehensive performance index, is selected through iteration. The exploration process includes setting the key parameters of the DE algorithm, such as population size NP, scaling factor F, and crossover coefficient CR, selecting sample points to construct an initial ship type parameter population, differential mutation and exchange of ship type parameter individuals within the population, fitness calculation and selection of ship type parameter individuals, and then executing this process a certain number of times as an iteration cycle until the new state generated by the algorithm exploration can meet the overall performance evaluation index.

[0064] This invention constructs an exploration space based on 28 parameters from the main dimensional parameters, hull line parameters, and functional layout parameters of the ship hull design, creating a 28-dimensional space (v1, v2…v…). n In this way, each individual in the population serves as a sample parameter for a ship-shaped design. Based on this design, a fitness function is set to evaluate each individual in differential evolution based on its overall performance, and individuals that do not meet the constant constraints are eliminated.

[0065] The population is initialized based on the original samples and the population size NP to obtain the initial population. Then, differential variation operations are performed on the individuals based on the scaling factor F and the crossover coefficient CR. i (g+1)=X r1 (g)+F(X r2 (g)-X r3 (g)), and crossover of the mutation results to enhance population diversity, the crossover results are as follows:

[0066]

[0067] Then, based on the overall performance of the ship type parameters, the fitness of each individual is calculated, and a greedy selection is performed to generate the new individuals:

[0068]

[0069] This results in a new population POP(t+1), which then enters a loop iteration to continuously optimize until a better conceptual solution set is output.

[0070] 3.3) Reasoning Stage

[0071] After the training phase, reinforcement learning can train a model capable of iteratively searching for combinations of hyperparameters for differential evolution. The hyperparameters represented by the model's output enable the differential evolution algorithm to find the conceptual design scheme with the best overall performance. For different inputs, the hyperparameter combinations that lead to the optimal conceptual design through differential evolution may differ. Therefore, the reinforcement learning model is essential for different conceptual design inputs, but it does not need to be retrained. For a set of conceptual design schemes, the model can quickly search for highly adaptive hyperparameter combinations using the learned strategy, and then optimize its multi-dimensional parameters using differential evolution.

[0072] 4) Use the optimized parameters to obtain the optimized differential evolution algorithm model;

[0073] 5) Use the optimized model to perform overall performance optimization on the main dimensions and hull lines data in the conceptual design scheme of the ship, and output the final conceptual scheme solution set.

[0074] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A rapid method for generating solution sets of ship conceptual design schemes based on multi-objective optimization, characterized in that, Includes the following steps: 1) Obtain the parameters of the conceptual design scheme and the range of values ​​for each parameter based on the requirements of the ship conceptual design scheme; The parameters include principal dimension parameters and profile parameters; 2) Establish a differential evolution optimization model and use differential evolution to optimize the ship's conceptual design scheme; The input to the differential evolution optimization model is the principal scale parameter and the profile data. Based on the hyperparameters represented by the next state, the model optimizes the current input conceptual design scheme and finally obtains the optimized conceptual design scheme. 3) Use reinforcement learning models to optimize the parameters of the differential evolution optimization model; the hyperparameters of differential evolution include: population size NP, scaling factor F, and crossover probability CR; The reinforcement learning model network structure includes an input layer, an output layer, and a 3-layer LSTM network structure. Each layer of the Long Short-Term Memory (LSTM) network has 35 neurons. The output layer is composed of a softmax function. There is a fully connected layer between the input layer, the output layer, and the 3-layer LSTM network structure. Optimizing the parameters of the differential evolution optimization model includes the following steps: 3.1) Initialize hyperparameters; 3.2) Determine the range of each of the three hyperparameters of differential evolution as the search space, and evaluate the reward value given by the environment for the action when choosing an action in the current state to reach the next state in the search space to update the three hyperparameters. 3.3) Optimize the current input concept design scheme based on the hyperparameters of the next state representation, and then input the next state into the reinforcement learning model for the next round of iteration to train the reinforcement learning model; 4) Use the optimized parameters to obtain the optimized differential evolution optimization model; After obtaining a set of hyperparameter combinations each time, calculate the overall performance evaluation value of the conceptual design output by the differential evolution optimization model under this set of hyperparameters, and select the hyperparameter combination with the largest overall performance evaluation value; 5) Use the optimized model to perform overall performance optimization on the main dimensions and hull lines data in the conceptual design scheme of the ship, and output the final conceptual scheme solution set.

Images