Sustainability performance assessment of sago industry supply chain using a multi-criteria adaptive fuzzy inference model

Related Works

Fuzzy inference systems (FIS) provide clear benefits over classical multicriteria decision-making (MCDM) methods, especially when it comes to managing the uncertainty and imprecision present in real-world decision-making situations (Yazdi et al., 2023). Recent studies have highlighted the effectiveness of fuzzy-based approaches in various applications, demonstrating their superiority in certain contexts (Yazdi et al., 2020) and also to deal with uncertain environments (Yazdi et al., 2022).

Several curious studies have been conducted on agri-food supply chain management, one of which is about supplier selection desire by integrating fuzzy into the Analytic Network Process to assess selection criteria. The study emphasizes a comprehensive approach for supply chain managers suggesting an integrated decision support framework and recommends further exploration of criteria relationships and further MCDM techniques (Ada, 2022).

The capacity of fuzzy systems to control ambiguity in decision-making is one of its main advantages (Sarfaraz et al., 2023). Traditional MCDM methods often rely on crisp values, which can be inadequate when dealing with ambiguous or imprecise data. For instance, utilized fuzzy PROMETHEE to navigate the complexities of treatment options for COVID-19, showcasing how fuzzy logic can clarify decision-making in uncertain environments (Yildirim et al., 2021). This aligns with findings by, who noted that many real-world scenarios involve uncertainty that crisp evaluations fail to capture, thus necessitating the integration of fuzzy evaluations into MCDM frameworks (Hinduja & Pandey, 2023).

Moreover, fuzzy systems enhance interpretability and flexibility in decision-making processes. The Fuzzy-ATOVIC method, for example, incorporates a FIS to improve the interpretability of the decision-making framework, allowing for a more nuanced understanding of the criteria and alternatives involved (Yusuf et al., 2022). This is particularly important in complex decision environments where stakeholders require clear justifications for choices made.

In addition to interpretability, fuzzy MCDM methods have been shown to produce more consistent and reliable results. demonstrated that integrating fuzzy sets of interval type-2 with the Best-Worst Method (BWM) yielded faster and more reliable outcomes compared to traditional MCDM methods, which often require extensive pairwise comparisons (Öztürk et al., 2022). This efficiency is crucial in time-sensitive decision-making scenarios, such as supplier selection or resource allocation.

Moreover, fuzzy MCDM approaches have been effectively used in a number of industries, such as manufacturing, civil engineering, and healthcare. For instance, a hybrid fuzzy MCDM approach was employed to optimize water supply planning, illustrating the versatility of fuzzy methods in addressing diverse decision-making challenges (Noori et al., 2020). Similar to this, fuzzy TOPSIS’s usefulness in handling the intricacies of real-world decision-making was shown when it was applied to the supplier selection process for construction projects (Zaman & Mishra, 2023).

The ANFIS is becoming known as a powerful modeling tool due to its ability to integrate the benefits of fuzzy logic systems and neural networks. ANFIS is commonly used because it has the ability to accurately model nonlinear systems. Research has shown that when it comes to forecasting outcomes in fields like agriculture and medicine, ANFIS performs better than other machine learning approaches like Support Vector Machines (SVM) and K-Nearest Neighbors (KNN) (Gökkuş et al., 2022). The researchers in the field of agriculture employed ANFIS to precisely predict grain yields, showcasing its efficacy in comprehending the intricacies of agricultural data (Gökkuş et al., 2022). ANFIS has been utilized in medical applications to forecast significant adverse cardiovascular events, demonstrating its resilience in managing medical data that contains intrinsic uncertainties (Karsidani et al., 2022). ANFIS demonstrates exceptional effectiveness in situations that involve inaccurate or partial data. The design of this system enables it to successfully handle uncertainty, a common difficulty encountered in real-world applications (Ibrahim, 2024). ANFIS’s versatility gives it an advantage over conventional statistical methods that may face difficulties in comparable circumstances.

ANFIS has a notable advantage in its capacity to integrate human-like reasoning using fuzzy logic, resulting in improved interpretability of the model’s judgments. This is especially advantageous in domains that need specialized expertise, such as environmental modeling and medical prognosis (Sulaiman et al., 2018). To summarize, ANFIS is preferred over other methods due to its precise modeling of nonlinear interactions, integration of human-like reasoning using fuzzy logic, better management of uncertainties, and improved performance through hybridization with optimization techniques.

Due to unsustainable practices, the growing demand for food has resulted in an increase in waste and environmental harm (Kumar et al., 2022a). Green supply chain management (GSCM), which emphasizes the integration of resource efficiency and environmental sustainability to improve performance across the supply chain, requires a thorough framework to evaluate performance. In addition, stakeholder collaboration is essential (Kazancoglu et al., 2018a). The sector must switch from a linear supply chain model to a circular one in order to maximize resource efficiency and minimize environmental effects. With this shift, the industry will receive socio-economic and environmental benefits (Lahane & Kant, 2021). The study by Kumar and Choubey (2023) highlights the significance of environmental variables and the necessity of routinely evaluating sustainability practices in order to ensure that they are in line with the SDGs. In the study, the multi-criteria fuzzy decision-making technique was applied.

Fuzzy Inference System (FIS)

Uncertain sets, or fuzzy sets, are collections of components with varying degrees of membership in 1965, Lotfi A. Zadeh separately proposed fuzzy sets, which extended the conventional concept of a set (Zadeh, 1965). In classical set theory, the evaluation of an element’s membership in a set is predicated on a bivalent condition, according to which an element can either be a member of the set or not. A membership function with a value in the real unit interval [0, 1] represents the iterative evaluation of an element’s membership within a set that is possible with fuzzy set theory. Fuzzy sets are special examples of indicator functions (also called characteristic functions) of classical sets since their membership functions only take values of 0 or 1. Fuzzy sets are hence a generalization of classical sets (Dubois, 1980). In fuzzy set theory, classic bivalent sets are often called crisp sets. Fuzzy set theory can be used in a number of domains, including bioinformatics, where data is imprecise or incomplete (Liang et al., 2006).

A fuzzy set is a pair (U, m) where m: U ➔ [0,1] is a membership function and U is a set (usually assumed to be non-empty). The universe of discourse is the reference set U (often represented by Ω or X), and the grade of membership of x in (U, m) is the value m(x) for each x ∈ U. The membership function of the fuzzy set A = (U, m) is defined as the function m = μ_A.

The fuzzy set (U, m) for a finite set U = {x₁, … ,x_n} is commonly represented by {m(x₁)/x₁, … ,m (x_n)/x_n}.

Assume x is in U. Then, x is classified as follows: fully included if m(x) = 1 (full member), partially included if 0 < m(x) < 1 (fuzzy member), and not included in the fuzzy set (U, m) if m(x) = 0 (no member).

The symbol SF(U) (or occasionally just F(U)) represents the (crisp) set of all fuzzy sets on a universe U (Beg & Ashraf, 2009).

The Fuzzy Inference System, according to Štěpnička & Mandal (2018), is a model framework that may be used to transform impression and uncertainty input based on fuzzy rules and principles into understandable output for decision makers. Fuzzy rules and principles consist of five parts: crisp output, defuzzification process, aggregation, fuzzification, fuzzy rules, and crisp input. The Mamdani inference model is used in this work to handle ambiguous data when evaluating supply chain sustainability.

The process of converting crisp input into linguistic input, such as “low,” “medium,” and “high,” depending on each input’s degree of membership function ascertained by its curve type, is known as fuzziness (Jamshidi et al., 2018). A set of rules known as fuzzy rules, or rule bases, are established in accordance with the input that has been specified during the fuzzification phase. Experts can determine the combination of rules that a fuzzy system will produce, and fuzzy rules work to predict the output that the system will produce. As suggested by Phillis & Kouikoglou (2009), fuzzy rules can be identified prior to the experts selecting the combination of fuzzy rules. Several rules are applied to the input that enters the fuzzy system, and an aggregation mechanism can be used to acquire the final output. The process of turning fuzzy quantities into clear output is called defuzzification.

Adaptive Neuro Fuzzy Inference System (ANFIS)

A hybrid of a fuzzy inference system (FIS) and an artificial neural network (ANN), Jang (1993) developed the Adaptive Neuro Fuzzy Inference System (ANFIS) model. The ANFIS model framework attempts to address the drawbacks of the ANN and FIS models so that the model can be trained and account for ambiguous and uncertain data. The ANFIS model framework’s architecture is depicted in Figure 1.

Figure 1. ANFIS structure.

Using language knowledge and numerical data for dynamic learning, an ANFIS is a hybrid intelligent system that integrates artificial neural networks and fuzzy logic systems to represent complicated, non-linear data dynamics. (Salleh & Hussain, 2016; Παπαγεωργίου et al., 2020).

The ANFIS process can be summarized by the following essential steps:

1. Fuzzification: The first step involves using membership functions (MFs) to convert exact input values into fuzzy values. These individuals define the manner in which each input variable is assigned to fuzzy sets. The selection of membership functions (MFs) is of utmost importance as they directly impact the system’s capacity to extrapolate from the training data (Chen et al., 2016; Salleh & Hussain, 2016). Multiple types of MFs can be utilized, such as triangular, trapezoidal, and Gaussian functions, according on the specific needs of the application (Chen et al., 2016). The ANFIS model at layer 1 can be seen in Equation 1.

2. Rule Base Formation: During this stage, a collection of imprecise rules is created using either expert knowledge or data-driven methods. The rules often adhere to the Takagi-Sugeno-Kang (TSK) model structure, which permits the use of linear output functions that are dependent on the fuzzy inputs (Chen et al., 2016). The number of rules can significantly affect how well the model performs. Insufficient rules may result in underfitting, while an excessive number might lead to overfitting (Baseri & Belali-Owsia, 2016).

Each node in this tier comprises a production T-norm operator, which functions as the node’s core component. Layer 2 in this system combines information transfer from layer 1 and replicates all incoming signals before sending them out as outputs. The result of the product layer is represented in Equation 2.

Each node in this layer stands for the rule’s strength. The weights in this layer control the output.

3. Inference Engine: The inference engine analyzes the fuzzy rules and merges the outcomes to get a fuzzy output. Fuzzy operators (AND, OR) are applied to the fuzzy sets provided by the rules in order to achieve this. The result of this stage is a fuzzy set that depicts the collective impacts of all rules (Salleh & Hussain, 2016). Layer 3 determines the normalization of the firing strength illustrated in Equation 3.

4. Defuzzification: The inference engine’s imprecise output needs to be converted into a precise value. To find the fuzzy set’s center of gravity, defuzzification techniques like the centroid approach are applied. This yields a single output value that can be comprehended in the context of the original problem (Chen et al., 2016; Salleh & Hussain, 2016). Layer 4 is an adaptive layer marked with a box diagram. Equation 4 illustrates how the parameters in this layer are derived from the subsequent parameters (p, q, and r).

where $\{p_i,q_i,r_i\}$ is the parameter set and $\overline{{\omega }_i}$ is the layer 3 output.

5. Learning Algorithm: The ANFIS system maximizes the parameters of the membership functions and, by extension, the parameters of the rules by employing a hybrid learning technique that combines gradient descent with least squares estimation. ANFIS can modify its structure by using the training data, which leads to an enhancement in its predicted accuracy as time progresses (Salleh & Hussain, 2016). The training phase is characterized by iteration, during which the system consistently improves its parameters until it reaches a suitable level of performance. Layer 5, the last layer in the ANFIS model, is represented by Equation 5, which is the model output that totals all of the model’s inputs.

6. Ultimately, the ANFIS model that has been trained is then evaluated using a distinct dataset in order to evaluate its ability to make accurate predictions. Ensuring that the model generalizes effectively to unfamiliar data is a critical step in machine learning applications, since it is a common difficulty (Baseri & Belali-Owsia, 2016; Παπαγεωργίου et al., 2020).

Research Framework and Stages

We employ a study methodology similar to that of Yani et al. (2022) to evaluate sustainability in the supply chain of the sugar agro-industry by varying the indicators in each dimension according to our research subjects, which serves to indicate the sustainability level. The research framework is divided into four stages: identification of sustainability indicators, FIS modelling, ANFIS modelling, and analysis of critical indicators. To assess each dimension’s performance, the FIS model was developed. In contrast, the value of each dimension was integrated into the overall supply chain sustainability performance using the ANFIS model. Figure 2 depicts the research framework.

Figure 2. Research framework.

The first step of the research is to establish indicators for each dimension. Next, the FIS model is developed to evaluate each dimension’s sustainability performance, and finally, the aggregation of the sustainability performance of each dimension with the ANFIS model. Additionally, important metrics for boosting the sago industry’s supply chain sustainability were examined. The stages of the research are shown in Figure 3. The following are the study’s steps:

Figure 3. Research stages.

Identification of sustainability indicators

Finding sustainability indicators for the case study on the sago agro-industry is the first step in this research. Field observations and a literature review were used to identify sustainability indicators, which were validated through in-depth interviews with practitioners and experts. The indicators that have been selected had to undergo the validation process and critique from relevant experts and stakeholders (Waas et al., 2014). Indicators in each dimension are needed to assess the sustainability level of each dimension. The indicator value of each dimension can be in the form of qualitative or quantitative data.

Sustainability indicators describe and provide information on the sustainability achievements of an organization in each dimension (Juwana et al., 2012). The metrics selected for evaluating supply chain sustainability must represent the degree of performance of the specified sustainability aspects and can be analyzed from a variety of angles, including social responsibility, environmental impact, and economic viability (Nijkamp and Vreeker, 2000; Waas et al., 2014). Sustainability indicators must be supported by qualitative or quantitative data (Galal and Moneim, 2016; Popovic et al., 2018).

Sustainability assessment with FIS model

A FIS is a system used to control how a system’s input and output variables relate to one another. Three categories exist for FIS: Tsukamoto, Sugeno, and Mamdani (Castillo et al., 2007). The primary distinction between Mamdani and Sugeno is the outcome of fuzzy rules. As a result of the rule, the Mamdani type employs a fuzzy set, whereas the Sugeno type employs a linear function. The result of each fuzzy rule for the Tsukamoto type employs a monotonic membership function. In this investigation, the Mamdani type was employed.

The FIS transforms crisp input into fuzzy input using a fuzzification interface. After fuzzification, a rule base is developed. Knowledge basis and rule base are interchangeable terms. Ultimately, the process of defuzzification is employed to transform the output’s fuzzy value into its crisp value.

The FIS model’s fuzzification stage represents sustainability indicator data for each dimension. Concerning the target of each indicator, all indicators are scaled with five linguistic levels: very low (VL), low (L), moderate (M), high (H), and very high (VH). If the target is minimal, the linguistic scale will be reversed; very high (VH) will start at the lowest value.

Using a specific membership function, the fuzzification stage converts observed data into fuzzy integers. The membership function in this study is a triangular fuzzy number (TFN). In order to produce consequences or outputs, FIS rules are arranged in the membership function according to the quantity of inputs and language levels. The rule number for each dimension is calculated by multiplying the linguistic levels in the membership function (m) by the number of indicators (n), as specified in Equation 6.

Each sustainability dimension has five indicators, and each indicator has five linguistic levels in the membership function, so 3125 rules on economic, social, and environmental aspects must be written for each FIS model.

The operator “AND” is associated with the FIS model’s input variables. The FIS model used the Mamdani model, and the output was represented by a linguistic label. Linguistic labels for sustainability performance are classified into five levels: very unsustainable, unsustainable, almost sustainable, sustainable, and very sustainable. Experts confirm that the fuzzy rule’s output accurately reflects the relationship between the input variables. This study employs approaches proposed by Phillis et al. (2011) to generate output on the fuzzy rules basis.

The fuzzy set with the linguistic scale is assigned an integer value of 1, 2 … so on, where one is associated with the least sustainable fuzzy set. The fuzzy set VL is assigned a value of 1, L a value of 2, M a value of 3, H a value of 4, and VH a value of 5. Furthermore, each input combination in each rule is summed to get the output value (OV), which is grouped into five linguistic scales.

Three thousand one hundred twenty-five rules represented the sustainability dimensions. For example, for rule number 2,675, social sustainability is sustainable if (institutional support is very high) and (local labor absorption is low) and (infrastructure to support activities is low) and (workers’ welfare is very high) and (participation of farmers in partnership is very high). The result is as follows:

The implication technique “min” and the aggregation method “max” are used by the FIS model for this study to assess how well the sustainability dimension is performing. The centroid defuzzification approach is applied. Table 1 provides a summary of the parameters that were utilized to create the FIS model and assess the sustainability dimension’s effectiveness

Aggregation of sustainability performance with the ANFIS model

The FIS and Artificial Neural Networks (ANN) learning benefits are combined in ANFIS. Despite system uncertainty, the FIS rule base describes the relationship between output and input parameters. ANN trains data and finds the best settings for the FIS membership function to get fuzzy rules and membership functions accordingly (Tozan and Vayvay, 2008). Using a FIS model, ANFIS is a straightforward data learning technique that transforms inputs into desired outputs. Fuzzification, product, normalization, defuzzification, and total output are the five primary process stages in ANFIS operations (Paul et al., 2015). The ANFIS model is structured similarly to the FIS model, except for estimating the membership function parameters and the FIS rules (Abdel-Aleem et al., 2017).

Several factors must be considered when developing the ANFIS model, including the variable input membership function, the number of training data pairs, epochs, and the model’s error tolerance. In ANFIS modelling, the membership functions (MFs) use grid partitioning by five linguistic MFs (very low, low, medium, high, and very high). Gaussian MFs were used in this model because they accurately describe the data distribution of a real-world problem (Moghaddamnia et al., 2009). The ANFIS learning algorithm combines the least squares estimator (LSE) and error backpropagation (EBP) methods. The model was trained for 10,000 epochs with zero error tolerance. A suitable ANFIS model should have a < 0.1 error (Sun et al., 2015). Table 2 lists the factors that were utilized in the development of the ANFIS model to aggregate the sustainability supply chain of the sago agro-industry.

Analysis of key indicators

The key indicators should be determined to design a strategic program to increase or maintain supply chain sustainability. Critical indicators are analyzed with the cosine amplitude method (CAM) at this stage. CAM is recommended because it effectively synthesizes fuzzy indicator measures (Ross, 2010). CAM looks for the similarity of data i and data j, symbolized by rij. Data i and j are n data, each having a membership function m in the fuzzy membership function. Every component of a relationship (r_ij) is the outcome of a pairwise comparison between two data samples (x_i and x_j, for example), where the relationship’s strength is expressed by the membership value. Like other similarity relations, the relation matrix will have n × n dimensions and be reflexive and symmetric, making it a tolerance relation. The score of the similarity between data i and j (r_ij) is formulated using Equation 7 and, like all similarity methods, guarantees that 0 ≤ r_ij ≤ 1 (Ross, 2010).

Next, the value of r_ij is arranged into a matrix, and then the average value is searched and normalized to determine the sensitivity value of each indicator. Finally, the value of the key indicators of each sustainability dimension can be determined.

Collecting Data

Specific sustainability indicators assess supply chain sustainability in the sago agro-industry. This study used 15 indicators with three sustainability dimensions, grouped as qualitative and quantitative data and analyzed using the formula in Table 3. Quantitative data on sustainability indicators were obtained through field surveys, in-depth interviews with supply chain actors, and previous research. Qualitative data was obtained by expert assessment and in-depth interviews with supply chain actors. Supply chain and sago sustainability experts include business actors, academics, and researchers.

This study generated 1,000 data sets using random numbers from zero to 100 by considering all the rules for the input-output relationship to develop the ANFIS model because of the limited availability of actual data. Seven hundred pairs were used for data training, and 300 pairs were used for data testing.

Verification and Validation Model

This study’s verification and validation model refers to Sargent (2013). The validation procedure ensured the accuracy of the formula and the model. At this point, literature studies back up the FIS model’s sustainability indicators. Model verification using FIS and ANFIS guarantees the lowest potential error in the model. Experts with knowledge of the model and current real-world conditions perform validation of conceptual models.

The conceptual model is validated by guaranteeing that literature studies justify the sustainability indicators. Fifteen sustainability indicators are used to assess supply chain sustainability performance. All sustainability indicators were determined using field surveys, expert opinions, interviews, and literature reviews. Table 4 shows evidence of verification of sustainability indicators based on a literature review.

By assessing the modeling performance of FIS and ANFIS, the FIS and ANFIS models are verified. RMSE, the number of membership functions, rules, and error values in training and testing are used in the evaluation of the FIS and ANFIS models.

One of the most widely used metrics for assessing continuous error models, RMSE, was used to evaluate the ANFIS model. RMSE is a calculation made to find the minor error from the following parameters in the forward step, fix the value of the premise parameter in the backward step with error propagation, and then calculate the error output from the network (Fatkhurrozi et al., 2012). RMSE stands for root mean square error, as the name would imply. RMSE is described in Equation 8 below. The number of data is n, x_i is the predicted value i to n, and y_i is the observed value i to n.

Sago Agro-industry Supply Chain and Sustainability Indicators Value

The goal of any supply chain is to maximize profits from the overall supply chain activity (Chopra and Meindl, 2013). How thriving operations are coordinated across supply chain levels to create value for consumers and enhance profits for each actor in the supply chain is a measure of the success of the supply chain (Somashekhar et al., 2014). The supply chain network’s borders are delineated by the supply chain structure, which also identifies the key participants and their respective responsibilities. The supply chain structure also describes all of the institutional arrangements, configurations, and other components that make up the network and support different business operations. In this research, the supply chain network boundary of the sago agro-industry consists of suppliers, manufacturers, and distributors, which are depicted in the supply chain structure in Figure 4.

Figure 4. Sago agro-industry supply chain configuration.

Industrial-scale sago agro-industry uses three sources of raw material: sago trunks from sago farmers or traders, sago trunks from factory gardens, and wet sago starch from small-scale wet sago mills. Meanwhile, the small-medium-scale sago agro-industry obtains raw materials only from sago farmers. Distributors will distribute sago starch to retailers or downstream industries.

Three criteria are used to evaluate the sustainability of the supply chain in the sago agroindustry: economic, social, and environmental. Based on a review of the literature, in-depth interviews, and expert validation, fifteen indicators have been chosen to evaluate environmental, social, and economic sustainability (Table 4).

Table 5 shows the target, maximum, and minimum indicator values for evaluating the sago industry’s supply chain sustainability performance. The proposed model is built with quantitative and qualitative data (Galal and Moneim, 2016; Popovic et al., 2018). Observations, interviews, and measurements were used to collect quantitative data, while expert judgment and a fuzzy membership function were used to collect qualitative data. Qualitative indicators are those that have linguistic labels.

The Performance of the Dimensions of Supply Chain Sustainability

The FIS model is used to evaluate the performance of each sustainability dimension using five indicators as inputs. Three models for evaluating the environmental, social, and economic aspects of sustainability were created using FIS. Because every dimension has five indicators, 3,125 rules (5 ⁵ in total) must be created to determine the sustainability performance of every dimension. Figure 5 depicts the fuzzy membership function of the FIS model’s input-output to determine the sustainability performance of the economic dimension.

Figure 5. Membership function of input and output for the economic dimension.

a) membership function of input 1. b) membership function of input 2. c) membership function of input 3. d) membership function of input 4. e) membership function of input 5. f) membership function of output.

The consequences of fuzzy rules in this study refer to the methodology proposed by Phillis et al. (2011). Figure 6 depicts the FIS framework for assigning supply chain sustainability performance for the social dimension, while Figure 7 depicts the rule display surfaces. Here are some examples of rules for each of the dimensions generated by the FIS:

Figure 6. FIS model framework for assessing social dimension sustainability.

Figure 7. The fuzzy rules surface.

The Mamdani type with a TFN membership function and a centroid defuzzification function was used to build the FIS model. Experts carry out the validation of the rule base in the FIS model. Following the generation of the FIS model, the data in Table 5 are used to assess the sustainability of each dimension in the small- to medium-sized industrial sago agri sector. The industrial-scale sago agro-industry supply chain’s sustainability was valued at 37.74, 55.29, and 53.17 in the economic, social, and environmental domains, respectively. In the meantime, the small- and medium-scale sago agro-industry supply chain’s sustainability scores in terms of economic, social, and environmental aspects were 56.61, 50, and 24.88, respectively. Figure 8 depicts the economic, social, and environmental sustainability of small- to medium-sized industrial sago agri sector supply chains.

Figure 8. Sustainability performance in each dimension.

Yusuf et al. (2019) assessed the sustainability of the sago agro-industry in South Sorong, Papua using the MDS technique with five dimensions and 25 indicators. The difference in indicators used, analytical methods, and research locations will, of course, result in different sustainability performances, so the results of this study cannot be compared. However, the similarity between this study and previous studies is that the value of the environmental dimension is lower than that of the economic and social dimensions. Previous research using the FIS model to assess supply chain sustainability has been carried out by Yani et al. (2022) in the sugarcane agro-industry with four dimensions and 24 indicators.

On the industrial scale, the lowest value is in the economic dimension, while on the small and medium scale, it is in the environmental dimension. The economic dimension indicators with the lowest value for the industrial scale are fair profit distribution among supply chain actors, a sufficient supply of raw materials to meet capacity, and a marketing network. The environmental dimension indicators with the lowest value for the small-medium scale are waste utilization and availability of waste management facilities.

Improving the sustainability performance of the supply chain in the sago agroindustry requires a close examination of supply chain risk and equitable profit distribution among supply chain participants, on both an industrial and a small-medium scale. Some literature states that agro-industry performance can be improved by efficient supply chain risk management (Safriyana et al., 2019; Septiani et al., 2016; Suripto et al., 2018; Zainuddin et al., 2017). In order to achieve sustainability through cooperation amongst supply chain participants, supply chain risk management employs methods, methodologies, and instruments to manage risk along the supply chain (Mulyati and Geldermann, 2017). One approach to optimizing supply chain risk and mitigation is acceptable risk and balanced profit distribution from the supply chain. It is important to acknowledge the significance of this approach in situations of uncertainty, particularly when working with multiple stakeholders, and to take into account the effects of past collaborative decision-making efforts. As a result, a fair and balanced risk and profit approach seeks to achieve a win-win solution by sharing and distributing profit and risk among all stakeholders to optimize supply chain performance (Chen, 2015; Palsule-Desai, 2013; Qian et al., 2013).

Important indicators in the social dimension are the participation of farmers in partnership and institutional support in both the small- to medium-sized industrial sago supply chains. Cooperation with suppliers has a positive impact on company performance. By reducing transaction costs and obtaining valuable technical resources, a comparative advantage in performance can be achieved through collaboration with suppliers (Wang and Dai, 2018). Previous researchers have mentioned institutional advantages for supply chains. Astuti et al. (2010) stated that the effectiveness and efficiency of the supply chain in achieving its goals are increased through the identification of needs and institutional structures in the supply chain. Institutions have a significant relationship with two elements, namely, information sharing and collaborative planning, which can increase profits and reduce transaction costs along the supply chain (Kalyar et al., 2013). Institutional is an effort to design interaction patterns so that they can carry out transactions between economic actors and various institutions, such as banks and governments (Alkadafi, 2014). The absence of institutions significantly impacts supply chain performance (Silvestre, 2015). Sustainable performance mitigates the consequences of the collaboration barriers required to establish sustainable supply chain productivity (Kumar and Goswami, 2019).

Important indicators in the environmental dimension are complaints about agro-industrial waste, waste utilization, and availability of waste management facilities. The complaints about agro-industrial waste indicator has the highest value of all the indicators in the environmental dimension. However, the availability of waste management facilities is deficient and even non-existent in the small- to medium-sized industrial sago agricultural sector. The community around the sago agro-industry that probably causes this are workers who depend on the sago agro-industry for their income. Some are concerned about waste and losing their source of income if the sago factory is closed.

Processing sago palms into sago starch requires water to extract the starch. The processing will produce both liquid and solid waste. Sago stalks are washed, hulled, and grated before being milled or processed for pulping during the process for extracting sago starch. As sago waste, a lot of water is released into the environment during production. The waste contains high levels of nitrogen compounds, cyano-glucosides, insoluble fiber, unextracted starch, fiber, suspended particles, and carbs. It contains high concentrations of organic materials, including lipids, carbohydrates, and proteins, is acidic, and emits a foul odor that causes pollution and worsens global environmental quality (Wee et al., 2017). The sago agro-industry, both industrial- and small-medium-scale, must pay attention to indicators L4 and L5 because the content of this waste can worsen environmental quality even though there are no complaints from the community about it.

The Overall Supply Chain Sustainability Performance

The aggregate values ought to depict the overall sustainability performance of the supply chain, which is obtained from the sustainability performance of the preceding stage’s dimension. Three parameters are input into the ANFIS approach, which aggregates sustainability performance for each measurement: the economic, social, and environmental dimensions. The sago agro-industry’s sustainable supply chain is the model’s result. The supply chain sustainability performance was determined for this study using the ANFIS model with grid partition initiation. The ANFIS model is structured using Gaussian MFs and three input variables, each with five MF levels, resulting in 125 rules that produce one output as the supply chain sustainability performance. The learning process uses 700 data sets and 300 data sets for testing. The data set was trained using a hybrid learning algorithm with 10000 epochs.

The production of 1000 data sets using random numbers ranging from zero to 100 prompts inquiries about the dependability of this method in real-world situations. Although the usage of random numbers is widespread in several domains, it is important to critically analyze the fundamental principles of randomness and how they affect the reliability of data in order to make accurate assumptions. The accuracy of generating random numbers is essential for ensuring the integrity of data. Pseudorandom number generators (PRNGs) produce random sequences using deterministic algorithms, which can lead to worries about their unpredictability and possible biases (Heese et al., 2021). While high-quality pseudorandom number generators (PRNGs) may be enough for machine learning tasks, any biases present in the PRNG can cause inconsistent outcomes in applications that require sensitivity (Aghamohammadi & Crutchfield, 2017). The effectiveness of random numbers is contingent upon the specific environment in which they are employed. Effective randomization techniques in clinical trials improve outcomes, however producing random numbers without taking into account the specific needs of the study can result in biased results (Pajcin et al., 2022). While constrained randomization enhances statistical features, the generation of random numbers does not ensure the creation of representative data sets (Peng et al., 2018).

It can be seen in Figure 9 that the error value significantly decreased at the 0-4000th epoch and resulted in the smallest error value at the 8300th epoch with a value of 0.007267. Table 6 shows a summary of the parameters and factors considered for developing the ANFIS model for the aggregation of the overall supply chain sustainability of the agro-industry of sago.

Figure 9. ANFIS training.

ANFIS training is critical for creating a model and determining the parameters with the least error. The ANFIS model’s training data have an error of 0.007267, while the test data have an error of 0.009306. The error condition for model training is < 0.1 (Sun et al., 2015). RMSE was also used to test the ANFIS model. With an RMSE value of 0.009592, the RMSE parameters for ANFIS performance evaluation show that the ANFIS model performs well. The RMSE value of 0.009592 indicates that the average predicted value differs from the actual value by 0.009592. The ANFIS model can handle the aggregation of the sago-agro-industry supply chain sustainability, according to the performance evaluation.

Figure 10 depicts the ANFIS model’s MFs following the training process. The MFs scale changes dynamically at each linguistic level, indicating that the model accurately reflects the training data. Figure 10 demonstrates the importance of training data in configuring MFs to predict the output. As a result, compared to the previously mentioned ANFIS model, the ANFIS model with grid partition can aggregate overall supply chain sustainability performance with near-zero errors.

Figure 10. Membership function of the ANFIS model before (a) and after (b) the training process.

Membership function of the ANFIS model before (a) and after (b) the training process.

ANFIS successfully designed three input dimensions of sustainability and one output from aggregating the total supply chain sustainability performance by launching the grid partition model. Figure 11 depicts the ANFIS model’s architecture. Finally, the ANFIS model is used to aggregate the overall sustainability performance of the supply chain. By utilizing the ANFIS model to aggregate the total supply chain sustainability performance, this study validates the performance of the sustainability component. The validation results indicate that the sustainability performance of the supply chain for the sago agroindustry is 44.25 at the industrial scale and 48.81 at the small- and medium-scale levels. The supply chains of both businesses fall within the almost sustainable group. Using an error value, the ANFIS model is verified in order to aggregate the supply chain sustainability performance overall.

Figure 11. ANFIS model architecture for aggregating supply chain sustainability performance.

The ANFIS model has also been used to assess country sustainability (Nilashi et al., 2018; Tan et al., 2017). The lowest error value attained throughout the training process is significantly influenced by the quantity of inputs, data sets, membership functions, and epochs.

The sustainability performance evaluation model’s benefit using the FIS and ANFIS techniques is that it can be used repeatedly to assess sustainability performance regularly by ensuring that indicators and ranges of values are still relevant. Managers must measure sustainability performance periodically to develop an action plan (Sharma et al., 2021). Apart from its advantages, the ANFIS ensemble certainly also has weaknesses, namely the large computational process and energy consumption (Tomasiello et al., 2023).

Analysis of Key Indicators

Key indicators are analyzed using the CAM approach to determine which ones can significantly improve sustainability performance. Key indicators need to be sought for each dimension of sustainability, economic, social, and environmental. CAM shows that indicators with high values are key indicators that should be used to improve sustainability performance.

The key performance metrics in supply chains for the small- and medium-sized sago agroindustry differ slightly. Fair profit distribution among supply chain actors is one of the primary indications of the economic dimension in the industrial-scale sago agro-industry supply chain (E2), differences in product demand (E4), and market access/marketing networks (E5). In the small-medium scale, they are E4, supply chain risk (E1), and E5. The critical indicator E2 is in line with previous research on the supply chain of the coffee, cocoa, and sugarcane agro-industry, which states that fair profit sharing, is the primary indicator of the economic dimension (Jaya et al., 2013; Sriwana et al., 2017; Yani et al., 2022). Figure 12 presents the key indicators of the economic dimension in the supply chain of the small- to medium-sized industrial sago sector.

Figure 12. Key indicators of the economic dimension.

The key indicators of the social dimension in the industrial-scale sago agro-industry supply chain include the availability of infrastructure to support activities (S3), institutional support for supply chains (S1), and workers’ welfare (S4). In the small-medium-scale industry, they are S4, S3, and S1. Although the values and order of the key indicators in the industrial scale and the small-medium scale are different, the key indicators of the social dimension in both industrial scales are the same, namely, S1, S3, and S4. Given that their low values might not even exist to enhance the sustainability performance of the supply chain for the sago agroindustry, the S1 indicators must be enhanced. The critical indicator S1 is in line with previous research on the cocoa and sugarcane agro-industry supply chain, which states that institutional support is the primary indicator of the social dimension (Sriwana et al., 2017; Yani et al., 2022). Figure 13 shows critical indicators of the social dimension in the supply chain of the medium- and small-scale Sago industry.

Figure 13. Key indicators of the social dimension.

The key indicators of the environmental dimension in the industrial-scale sago agro-industry supply chain are fuel consumption emissions (L1), availability of waste management facilities (L5), and water consumption (L2). In the small-medium scale, they are L2, waste utilization (reuse & recycle) (L4), and L1. Critical to the assessment of the environmental dimension’s sustainability are all of the indicators employed. The indicators that still need to be improved by the two scales of the sago industry are L4 and L5, which are currently very minimal and have not even been carried out in the small-medium scale sago agro-industry. Previous research on the coffee and sugarcane agro-industries supply chain has also stated that it is a significant indicator of the environmental dimension and needs more attention (Jaya et al., 2013; Yani et al., 2022). Figure 14 shows the critical indicators of ecological measurements in the supply chain of the small- to medium-sized industrial sago agricultural sector.

Figure 14. Key indicators of the environmental dimension.

The results of the analysis of key indicators are the critical indicators that, if the value is increased, will significantly improve sustainability performance. Analysis of these key indicators is needed to formulate strategic and action plans to improve sustainability performance. Fair profit sharing among supply chain participants, institutional support for supply chains, waste utilization (reuse & recycle), and the availability of waste management facilities are important factors for enhancing sustainability performance in each dimension.

Managerial implication

The sago agro-industry’s sustainability performance evaluation model was created with the use of carefully selected, industry-relevant indicators. All parties involved in the supply chain must work together to ensure the sustainability of the sago agro-industry supply chain. To find out the sustainability status, managers at focused companies can periodically assess the sustainability performance of the supply chain. The advantage of the sustainability performance assessment model using FIS and ANFIS is that it can be used frequently and updated regularly. Identifying critical indicators helps managers develop strategic plans and practices to improve sustainability performance.

The methodology this research provides can be used to other agro-industrial sectors that are confronting comparable sustainability issues, meaning that its ramifications reach beyond the sago business. The model’s emphasis on multi-criteria evaluation and adaptability positions it as a valuable tool for practitioners and policymakers aiming to enhance sustainability performance across various supply chains.

Limitation and Future Research

Three distinct dimensions—economic, social, and environmental—are the only ones included in the model this study suggests for assessing the sustainability of the supply chain for the sago agroindustry. Other study may incorporate other aspects, such as resources or technology. Due to limited data availability, only five indicators are utilized in each dimension. By changing the indicators and assessing the suitability and accessibility of the data for the study topic, the proposed model can be expanded to include a wider range of agro-industries. Equitable allocation of profits among participants in the supply chain is a crucial measure of the economic aspect, and additional investigation into the appropriate profit distribution model is required. Furthermore, it is necessary to do research on enhancing the institutional capacity of the sago agro-industry supply chain in order to address the crucial factors in the social aspect. It is crucial to develop a waste management strategy that includes reusing and recycling materials. Additionally, it is critical to determine the necessary facilities based on their environmental implications and associated costs.