3.2. General Framework
The flowchart of the MAPSO algorithm is illustrated in
Figure 1. Initially, parameters such as the number of ground stations, satellites, and the population size are input. A set of uniformly distributed weight vectors [
21] is then generated. The population
is initialized by generating a group of randomly distributed particles in space with initial velocities
(where
). An initial archive
is established to store the non-dominated solutions of population
, which is updated based on the clone strategy.
After completing the population initialization, we employ the archive update strategy to update the external archive. At the start of each iteration, the contents of archive are copied to prevent the loss of solution sets. The PSO search strategy selects leaders from the archive to update the population. After updating the population, the archive update strategy is employed again to update the sets and . The procedure is repeated until the end condition is met. The algorithm comprises three key components: the clone, the PSO search process, and the archive update operation. The upcoming sections will offer a detailed introduction to these three components.
3.3. Clone
As shown in Algorithm 1, the clone steps aim to enhance the diversity and convergence of the external archive
, thereby accelerating the PSO convergence speed, particularly for the parameter-independent MAPSO algorithm. The process involves the following steps: initialization, SDE distance calculation, selection and duplication, and generation of offspring population. The clone process is illustrated in
Figure 2. Below are the detailed descriptions.
Initialization: First, copy the external archive to a temporary archive and initialize a temporary archive as empty. Make sure that Archive remains intact in the following steps.
Shift-based density estimation (SDE) distance calculation: SDE is an improved density estimation strategy designed to address the challenges faced by Pareto-based evolutionary multi-objective optimization algorithms when handling multi-objective optimization problems. Traditional density estimation methods only reflect the distribution of individuals in the population, while SDE considers individuals’ distribution and convergence information. Calculate the SDE distance for each solution
in
The SDE distance measures the diversity of the solutions. Refer to the relevant literature for specific calculation steps and equations [
22].
Algorithm 1 Clone Process |
Input: |
Output: |
// Sort B based on the SDE metric if then end if for do Calculate Create solutions for and append to end for S←modify(C) Return S
|
Selection and duplication: Next, select
solutions from the temporary archive
for duplication. The number of duplications
for each solution
is calculated using the following (14):
where
presents the
SDE distance of the solution
, and
is the total number of solutions to be duplicated. This step ensures that solutions with more considerable SDE distances obtain more duplication opportunities, thereby enhancing the diversity and convergence of the population.
Generation of offspring population: Each solution
is duplicated
times and added to the temporary archive
, forming a new offspring population
. This population then undergoes evolutionary search, applying operations such as crossover and mutation, resulting in a new offspring population
. The clone selects superior solutions based on SDE distance, ensuring that the external archive
maintains good diversity and convergence, effectively guiding the evolution of population
in PSO and avoiding a local optimum. The evolutionary search strategy in MAPSO is implemented in the same manner as in NSGA-II [
23].
The clone guarantees the population’s diversity and convergence and accelerates MAPSO convergence speed. Through these steps, the external archive exhibits excellent diversity and convergence, promoting the practical evolution of the MAPSO population , making the algorithm more efficient in handling multi-objective optimization problems.
3.4. Archive Update
The archive update step in Algorithm 2 determines which solutions will be added to archive
. The performance of multi-objective algorithms is closely tied to selecting the external archive set. Therefore, choosing solutions that exhibit good convergence and diversity is crucial. The clone process is illustrated in
Figure 3. Below are the detailed descriptions.
Algorithm 2 Archive Update Process |
|
|
end if Return A
|
The archive update operation updates the external archive A through the following steps:
Merge archives: Combine archive with population , creating a joint population that includes both and .
Normalization: Normalize each solution in the joint population , standardizing the objective values to a range between 0 and 1.
Calculation fitness: Calculate the fitness value for each solution in
using Equation (15):
Non-dominated sorting: Pareto dominance filters out non-dominated solutions from C, forming the non-dominated solution set Cnd.
If the number of solutions in Cnd exceeds the population size , use the minimum angle selection strategy to select solutions added to ;
If the number of solutions in Cnd is less than or equal to N, directly add all solutions in Cnd to . Return updated archive: return the updated archive .
The is the fitness value of the solution in the joint population .
Fitness value serves as a standard for evaluating the convergence of the solutions. Subsequently, the non-dominated solution set is selected from -based Pareto dominance criteria.
Because in MAPSO, most solutions are non-dominated, the minimum angle selection strategy is adopted when the Pareto dominance criterion is no longer adequate. The steps of the minimum angle selection strategy, as detailed in Algorithm 3, can be divided into three main parts:
The angle calculation method directly affects particle distribution and search space exploration. This study uses the following (16) equation for calculating the angle between particles:
instead of (17):
The benefits of using Equation (16) (based on arcsin) are as follows:
- (1)
Intuitive geometric interpretation and visualization: The range of arcsinX is closer to actual movement angles, making it easier to interpret and visualize particle movement paths. Due to its smaller range, arcsinX can better describe angle changes in physical systems, preventing significant angle jumps and facilitating fine-tuning.
Assume a satellite is orbiting along a circular path around the Earth. The angle
between the satellite and a point on the Earth’s surface can be represented as
θ = arcsin(d/h), where
is the distance along the Earth’s surface from the point directly beneath the satellite to the point of interest, and
is the altitude of the satellite.
Algorithm 3 Minimum Angle Selection Process |
|
We compare it to a linear approximation of the actual movement angles to show that arcsinX provides a closer approximation. The linear approximation can be written as , which is only accurate for tiny angles ().
When we expand
with the Taylor series, we can use the general form of the Taylor series
using Equation (18).
For , we consider as Thus, the expansion becomes .
This expansion shows that for small values of , the linear approximation is reasonable. However, as increases, higher-order terms become significant, making the approximation of more accurate.
Figure 4 illustrates a comparison between the calculated actual movement angles
using the
function and a straightforward linear approximation
. The plot demonstrates that the arcsin function (blue curve) closely follows the accurate movement angles, especially for larger values of d, illustrating its superior accuracy. In contrast, the linear approximation (red dashed line) deviates more as d increases, highlighting the effectiveness of the arcsin approach in accurately representing movement angles.
- (2)
Numerical stability: varies more smoothly within its range, and its derivative does not tend to infinity, resulting in more stable numerical computations. In contrast, has larger derivative changes near boundary values, which can cause numerical instability.
- (3)
Fine-grained search: For multi-objective optimization problems in satellite data transmission, optimizing transmission paths and resource allocation requires high precision and fine adjustments. offers finer search granularity, ensuring better solutions even under complex constraints. The smooth and gradual changes of lead to more uniform particle distribution and stable computations.
In summary, Equation (16) based on is more suitable for PSO algorithms dealing with complex multi-objective optimization problems, such as satellite data transmission systems, due to its intuitive geometric interpretation, numerical stability, and fine-grained search capability.
- 3.
Add Solutions with Maximum Vector Angles:
The process of the minimum angle selection operation is illustrated in
Figure 5. The triangular point (▲) represents solutions in
R, while solid (•) belongs to
.
R represents the set of solutions that were removed in step (2).
well-distributed solution
is removed, indicating that some well-distributed solutions might be eliminated. MAPSO allows solutions removed from
R to be re-added to
if the number of solutions in
is less than N−m. Find the solution in
with the maximum vector angle relative to solutions in
using the following (19):
Add this solution to and remove it from . Repeat this process until the number of solutions in reaches . Merge and , removing solutions already in to form the new combined population . Return the updated archive .
3.5. PSO Search
The procedure for PSO is detailed in Algorithm 4. The PSO method is employed to update the population
. Before updating the solutions’ position information, it is crucial to determine
pbest and
gbest. In multi-objective optimization problems, solutions cannot be directly compared using a single objective value; however, Pareto dominance can be utilized. In contrast, for many-objective optimization problems, most solutions are non-dominated, which complicates the selection of suitable solutions (
pbest and
gbest) for updating the population. Therefore,
pbest and
gbest are chosen based on the particle angle size. Similar to conventional PSO, this process includes selecting
pbest and
gbest and updating the velocity and position (lines 2–3 of Algorithm 4). The specific steps are as follows:
Algorithm 4 PSO Algorithm |
|
|
|
Choosing
pbest and
gbest:
penalty factor
balances population diversity and convergence when choosing
pbest and
gbest.
higher
encourages diversity, while a lower
favors convergence. The equation for calculating
is given in Equation (20):
represents a scaling parameter, and denotes the vector angle between and , with N as the population size. This vector angle is normalized to range between 0° and 90°.
Figure 6 illustrates the selection of
. The red lines represent the reference vectors
and
.
smaller vector angle between
and
indicates that
has good diversity. By using the equation above to calculate
, a particle with a smaller angle can be chosen as
pbest1 to guide
rather than another particle with better diversity but worse convergence. Similarly,
should be selected as
gbest2 for its diversity rather than
, which might have better convergence. For MAOP, there is no single optimal solution, so a solution close to
from
A is selected as
to improve diversity. In
Figure 6,
and
will be chosen as
gbest1 and
gbest2, respectively. Sometimes,
and
are equal, indicating that the solution can provide diversity and convergence directions.
Updating particle position and velocity: Once
pbest and
gbest have been identified, the particle’s velocity
is determined using the following Equation (21):
where
represents the position of the
i-th solution in the population,
and
are evenly dispersed within the range of [0, 1],
denotes the inertia weight, and
c represents the learning factor.
balances the search direction between
pbesti and
gbesti, calculated as Equation (22):
Based on Equation (22), the overall search direction is primarily influenced by
pbesti and
gbesti. The branch search direction of
is governed by
, which emphasizes diversity rather than convergence. If
fitness is higher than
it indicates that
is closer to the Pareto front (PF) than
. The search direction is from
to
otherwise, it is the opposite. Finally, the position
is updated as follows (23):
If the position exceeds the boundaries in some dimensions, it is set to the boundary value, and the corresponding velocity is reset to zero.