1. Introduction
The Artificial Neural Network (ANN) has become the standard approach to solve a wide range of optimization problems due to the structural concepts of simulating the neurons’ activity of the human brain. The solid feature of ANN is mimic ability on how the brain works with the hope that we can build a system or a model that can produce collective decision making. As a result, AI practitioners heavily rely on the features of ANN which results in a lack of in-depth understanding of the operations conducted by ANN models. According to [
1], ANN is considered a non-interpretable model due to its “black-box” nature, which fails to determine the direction and magnitude of the neurons in the network. In addition, the regularization and training time of the ANN model is difficult to be set as a constant. Hence, Wan Abdullah [
2] embedded the Satisfiability (SAT) logical structure in a class of feedback ANN model, the Hopfield Neural Network (HNN). The role of SAT in HNN is to represent the neurons as meaningful indicators (or meaningless); when using such auxiliary inputs, one can use it as the basis of reasoning to explain the predicted outputs. In layman’s terms, SAT is the symbolic instruction to navigate the operations of HNN. The work capitalized the structure of SAT by proposing logical inference through the minimization of Horn Satisfiability (HornSAT) in HNN. The approach is capable to yield connection strengths (synaptic weight) between each neuron, resulting in an energy-minimizing dynamic network. However, HornSAT consists of redundant variables which deteriorate the practicality in representing real-life variables. Therefore, ref. [
3] incorporates another variant of SAT in HNN called 2 Satisfiability (2SAT). 2SAT is a class of systematic SAT whereby the number of variables in each clause is restricted to only
k variables, in this case,
k = 2. The performance of 2SAT in HNN can achieve 90% production of global minimum solutions, which indicates the credibility of 2SAT in representing the behavior of neurons in HNN.
Alternatively, ref. [
4] counters the problem with the existing SAT structure by proposing Random 2 Satisfiability (RAN2SAT) in the same network. RAN2SAT is the non-systematic variant of SAT which considers a non-restrictive number of variables in each clause, whereby RAN2SAT consists of clauses with
k = 1, 2. Such formulation is considered to promote logical variation, which is believed to widen the search space of global minimum solutions. The work can locate maximum production of global solutions which indicates RAN2SAT is successfully embedded in the operations of HNN. Despite acceptable results, little attention has been given to RAN2SAT that claimed has potential as optimal SAT in HNN. The repetitive final neuron state which results in overfitting solutions is still an issue to be solved. Recently, ref. [
5] proposed a new novel of non-systematic SAT named Major 2 Satisfiability (MAJ2SAT). The user interface provided in MAJ2SAT controls the structure of the logic where 2SAT is a dominant clause. As a result, MAJ2SAT can increase the neuron variation as the ratio of 2SAT increases. Although a restrictive training environment, the synaptic weight attained leads to global minima solutions. However, overfitting occurs when the 2SAT ratio is increased, which explains the creation of a repeated pattern of neurons states that disrupted the variation process and may reduce the precision of HNN. Thus far, there has been no investigation to seek a new novel of SAT that can provide non-overfitting solutions in HNN.
A metaheuristic is an effective method for obtaining an approximately optimal solution for the problems. For decades, a variety of metaheuristics with different classes was proposed and improved. The work by Zamani [
6] is worth exploring where it proposed a new metaheuristic inspired by migratory birds’ behavior during their long-distance aerial migrations named QANA. The features of QANA are long-term and short-term memories, a V-echelon communication topology, and quantum-based navigation with two mutation strategies and a qubit-crossover operator. These features enhance the population diversity and avoid premature convergence. The effectiveness of QANA was evaluated using benchmark functions and four engineering functions. QANA can outperform state-of-the-art algorithms. However, the work shows the significant findings of QANA in solving single and continuous problems. Another interesting work that is worth exploring is by [
7] improved Moth Flame Optimization (MFO) by introducing a migration strategy to enhance the exploration and maintain diversity. The performance of M-MFO was evaluated by conducting benchmark functions and compared with seven variants of MFO and eight state-of-the-art algorithms. It is worth mentioning that M-MFO managed to maintain while the size of the problem increased. Unfortunately, the work did not investigate the performance of M-MFO in the lowest population diversity. In 2019, Kasihmuddin [
8] implemented Estimation of Distribution in the DHNN to predict the possible final neuron states with the lowest minimum energy which leads to a global minimum solution. EDA is able to provide superior performance in all the metrics. Despite successful results, this work did not analyze the proposed model with different logical rules. In another development, Sathasivam [
9] was the first person to incorporate non-systematic SAT with metaheuristic. The work implementing Election Algorithm (EA) as a training algorithm in DHNN to optimize the structure of RAN2SAT. This work demonstrated EA and RAN2SAT’s capability to promote more diverse interpretations that lead to global minima solutions. Additionally, the EA can outperform another state-of-the-art algorithm due to its effective partitioning of solution space that reduces the complexity. However, the work did not analyze the impact of the retrieval phase. Note that metaheuristics allow us to effectively obtain approximate solutions to problems. However, extensive investigation is required to determine suitable metaheuristics for the specific problem.
Various social creatures exist in nature, such as ants, bees, and fish, where individuals from a given population cannot survive without the other members of that population. However, when we analyze the entire population, we see that the individual is extraordinarily intelligent through cooperation. Consequently, researchers have developed a type of algorithm to solve optimization problems that are inspired by these natures, named swarm intelligence algorithm (SI). One of the most widely used SI algorithms in the studies is
ABC, which simulates the foraging behavior of honeybees, initially proposed to optimize continuous optimization problems [
10]. The work was extended by [
11] that developed binary variants of
ABC named bitABC. The model uses the same framework as the original
ABC but uses bitwise operations to optimize the food source. A comprehensive comparison has been made with another variant binary
ABC with 13 selected benchmarks problem. The performance of BitABC is crucial in terms of final solution accuracy, convergence speed, and robustness. Subsequently, Kasihmuddin [
3] proposed
ABC as a training algorithm in HNN with 2SAT. This work demonstrates that
ABC is more competitive in HNN compared to the standard method in obtaining satisfactiry interpretation of 2SAT. However, the work lacks variation and diversification in the quality of the retrieval final neuron states. In another development,
ABC was incorporated with Double Layer Neural Network (DLNN) to solve the Bi-Level Programming Problem (BLPP) [
12].
ABC was improved by cooperating with two major components of the Genetic Algorithm (GA) which are crossover and mutation. The improved-
ABC is able to yield a high quality of global optimal solution with higher accuracy in a smaller time. The work proves that
ABC is flexible to co-operate with another metaheuristic. Currently, ref. [
13] proposed modification of solution update rule of basic
ABC replaced with a xor logic gate in binABC. The comprehensive comparison was conducted with six variants of binary
ABC by solving a modern benchmark problem (CEC2015). binABC is compatible in solving discrete problems. This finding shows that the update rule equation plays an important role in binary
ABC. However, there is no investigation on the performance of different logic gate operators in the update rule equation. Although many researchers show the stellar performance of binary
ABC in solving a variety of problems, there has been much uncertainty of which basic logic gate operators work well in solving discrete optimization problems.
Table 1 below summarizes the related research.
The important contributions of this study are listed as follows:
- (i)
This study introduced an additional layer of DHNN to generate a non-systematic logical structure with a consideration ratio of negative literals. The higher number of negative literals in the logical structure is considered to enhance diversified final neuron states retrieved.
- (ii)
This study explores the capability of the Artificial Bee Colony algorithm (ABC) with different logic gate operators in the update rule equation. The main purpose of ABC in this study is to find the correct logical structure according to a predetermined ratio of negative literal. A comprehensive comparison was conducted to reveal the effectiveness of ABC with the NAND operator. Next, the performance of ABC will be compared with three state-of-the-art metaheuristics. Statistical analysis was conducted to show that ABC is compatible in finding the correct logical structure.
The major contributions of the present study are as follows:
- (i)
A new novel of non-systematic Weighted Random k Satisfiability (rkSAT) is proposed by combining the first and second-order Satisfiability logical rule with consideration of negative literals.
- (ii)
The implementation of Artificial Bee Colony in the logic phase to control the distribution of negative literals in the logical structure.
- (iii)
A comprehensive comparative analysis of five different logic gate operators in the food source equation in terms of producing a correct number of negative literals for a higher ratio.
- (iv)
The proposed Artificial Bee Colony in the logic phase will be compared with the state-of-the-art binary algorithm.
Figure 1 illustrates the general flow of the proposed study to ensure the readers gain a better understanding of this approach. In the first place, the initialization of the logic phase occurs after the ratio of negative literal and number of literal is predetermined.
ABC will be utilized in the logic phase as the search technique in order to generate the correct structure of Weighted Random 2 Satisfiability (
). Note that the objective of the logic phase is to distribute a correct number of negative literal in the
. Then, the first phase of DHNN will optimize
to ensure the correct synaptic weight is obtained. The computation of the local field of the neuron state and the final energy happens in the testing phase. The quality of the output will be evaluated by comparing the final energy. Supposing the output achieved lies within the tolerance value, the final neuron states will achieve global minima solution, otherwise be trapped in local minima solution.
The outline of the current study includes
Section 2, provides the preliminary explanation on the Weighted Random
k Satisfiability.
Section 3 briefly explains Weighted Random 2 Satisfiability study in DHNN. Then, the proposed binary Artificial Bee Colony and other binary algorithms are discussed in
Section 4. The methods and experimental setup are given in
Section 5. The simulation of the study is discussed in
Section 6. Finally, concluding remarks are given in the last section.
2. Weighted Random k Satisfiability
Satisfiability (SAT) is a Boolean formula that is said to be satisfiable if satisfied assignments of the literals exist, which makes the formula true. SAT is categorized as NP-complete [
14] which inspires logic mining practitioners to utilize SAT as a form of symbolic instruction in representing the entries of a real dataset. SAT can be divided into two main structural classes which are systematic and non-systematic logical rule [
15]. According to the findings in [
4], non-systematic SAT provides more logical variation in terms of final neuron states. In this article, Weighted Random
k Satisfiability is introduced as a new class of non-systematic SAT with consideration of the desired number of negative literals. The number of negative literals in the SAT can be predetermined with a randomized choice of literals that can be either positive or negative. The general equation for Weighted
k Satisfiability can be formulated as follows:
whereby the value of 2 (second) and 1 (first) depicts the order of clauses
in
where
. Note that
and
denote the total number of second- and first-order clauses, respectively. The total combination of clauses in
can also be defined as
where
. Following Equation (1), the combination of orders in the formulation makes
a class of non-systematic SAT whereby the number of literals in each clause is not restricted to only one value of
k. Note that
and
are possible clause for
formulation. In other words,
has non-redundant literals, whereby
for any orders of
. By considering both order clauses, the total number of literal exists in
can be calculated by using Equation (2) as below:
According to the study by [
8], the assignments that correspond to
can be presented as
. The main goal of the
is to find a set of assignments that make the whole formulation become TRUE. In comparison with the previous structure such as Karim et al. [
15] and Alway et al. [
5], the main structural comparison of
is the feature of controlling the distribution of negative (negation) literals in the logical structure. Thus, a systematic mechanism is introduced to only allow a specific total number of negative literals in
.
is defined as the desired number of negative literals in
which can be calculated by using Equation (3) below:
is the ratio of negative literals existing in with a range of . The pre-determined value of depicts the percentage of negative literals, whereby when , the percentage of negative literals in is 50% out of all . The proposed step size in this article is . Conjointly, the value of can also is defined as , as the number of negative literals that should be considered must be less than or equal to a given number in the form of . For example, for , the value of is given by . Thus, using the value of , one possible negative literal of is given by . This indicates that a specific method is required to filter the type of logical rule that satisfies the condition in Equation (3).
Unfortunately, without an appropriate mechanism, the distribution of negative literal in
will potentially be skewed to one side or biased to a specific type of clause. Therefore, in this article, an additional optimization layer of controlling the distribution of the negative literal in
is presented. The value of
is set as the objective function with randomized literals selection which will be determined as negative or positive. In this context, a minimization task is introduced before generating the right structure of
which will be formulated based on Equation (4).
whereby the optimal value of the best fitness,
is equal to 0. Note that
is the total number of weights in
which can be formulated in Equation (5) as follows:
As
is the number of negative literals which should be achieved in order to satisfy Equation (6),
always accounts for a non-zero and non-negative value. Equation (6) presents the possible value of the weight,
:
where
can be any arbitrary negative literals in
k order clauses. Thus, to guarantee only the desired number of negative literals is achieved, without any bias towards a specific order of clauses, an additional phase named the logic phase is added to the model. This effort is to ensure
is optimally generated with respect to
. The role of the logic phase is to effectively generate
with the dynamic value of
and
by successfully minimizing Equation (4). In summary, the structural feature of
formula are presented as follows:
The structure combines a different order of clauses, with logical AND ().
The variable in the considers non-redundant with randomized selection to be negative or positive.
Consideration of the desired number of negative literals without any biased to a specific order of clauses.
Consecutively, the minimization task in the logic phase will be carried out using an optimization algorithm which will be explained in the next section.
3. Discrete Hopfield Neural Network
DHNN consists of interconnected neurons containing input and output layers but without hidden layers. Several features of DHNN include associative memory, fault tolerance, and energy minimization as the neuron state changes. Generally, the neurons in DHNN will be updated asynchronously until the final neuron state reaches the equilibrium state which corresponds to the solution of the optimization problem. Thus, the general formulation of updating rule of
-th neuron
in DHNN is given in Equation (7) as follows:
where
and
denote the synaptic weight and threshold of the network. In order to ensure the energy of DHNN decreases monotonically, we set
which demonstrates effective neuron state classification. The
will be embedded by assigning each variable with neurons to the defined cost function,
. Equations (8) and (9) below represent the generalized cost function of
:
Note that the value of
is proportional to the number of inconsistencies of the clauses [
4]. In this context,
will increase as the number of unsatisfied clauses increases. The lowest possible value that can be obtained for the cost function is
. In order to ensure the minimization of
, the following local field is utilized by using Equation (10):
is an updated neuron state given by Equation (11):
where
and
are second and first order synaptic weights, respectively. Hyperbolic Tangent Activation Function (HTAF) is utilized in Equation (11) to avoid the neuron state being updated linearly. In terms of neuron connection, synaptic weights can be obtained by using the Wan Abdullah method [
2] which compares Equation (8) and Equation (12) as long as the proposed
has at least one satisfied interpretation. The Lyapunov energy function that corresponds to the
logic is given as in Equation (12):
The value of
indicates the scalar quantity of the final neuron state which corresponds to the interpretation that models
. Interestingly, Equation (12) informs the network whether the final neuron state reaches the absolute minimum energy that corresponds to the optimal final states for
.
Figure 2 shows the schematic diagram for DHNN-
r2SAT. Note that the blue dotted lines represent the second-order clauses with four possibilities of neurons in the second-order clauses. Meanwhile, the green dotted lines represent first-order clauses with two possibilities of neurons.
7. Conclusions
The first contribution of this study introduces a systematic method in generating the logic with consideration of negative literal via Weighted Random 2 Satisfiability. The study shows that the logic phase is efficient to diversify the logical rule by controlling the dynamic of negative literal. The second contribution is that binary
ABC was utilized in the logic phase to find the correct structure of
r2SAT. Then, the performance of the proposed model was successfully compared with four variants of binary
ABC with different logic gate operators in the food source equation. The experiments show the compatibility of
in the logic phase by generating a higher number of negative literals. Then, the proposed model is compared with three state-of-the-art algorithms:
,
, and
.
ABC gives a solid performance as it requires few iterations to obtain the desired structure of logic. This study highlighted the capability of
in logic phase in generating the correct number of negative literals in a logical structure. In addition, this study also highlighted the capability of different logic gate operators in generating desired logical structure. The study can be further extended by considering the lowest values of
r which are 0.1, 0.2, and 0.3. Furthermore, different types of logic can generate in the logic phase, such as Random 2 Satisfiability [
4], High order of Random
k Satisfiability [
15], and Major 2 Satisfiability [
5]. The robust architecture of
r2SAT, integrated with an optimal operator for
ABC, provides a view on the possible random dynamics for the application of real-life bioinformatics problem. For example, the proposed
can be embedded into logic mining which extract the best logical rule that classifies single-nucleotide polymorphisms (SNPs) inside known genes associated with Alzheimer’s disease. This can lead to large-scale logic mining design incorporated with
, which has the ability to classify and predict.