Performance prediction of a textile reverse logistics system using DEA and ANFIS hybrid models

Abstract

Due to the growing call to embrace environmentally responsible and sustainable business practices, textile reverse logistics (TRL) and recovery practices, such as reusing, remanufacturing, or recycling, are gaining prominence. Textile recycling companies can simultaneously obtain economic and environmental benefits via more efficient RL practices. However, a system for measuring these efficiencies is paramount, as it is impossible to run a reverse logistics system efficiently without the ability to measure its performance. Studies on performance measurement of TRL firms are completely lacking, and those of the general RL literature use manual systems that require longer time and participation of many workers to complete. In this study, we develop a performance prediction model based on DEA and ANFIS. Data for the ANFIS were derived from the DEA computation. To enhance the model, PSO, GA, and Jaya algorithms were introduced to tweak the ANFIS parameters. Results from the ANFIS hybrid models reveal ANFIS-Jaya to have a better prediction accuracy with R² of 0.9832 and 0.9851 in training and testing datasets, respectively. This study contributes to the RL performance management literature and the limited research on used clothing collection, textile recycling, and RL performance management measurement.

Keywords

Textile reverse logistics DEA-ANFIS recycling Jaya algorithm

1 Introduction

Fast-paced fashion cycles and customers’ lack of commitment to sustainable purchasing patterns define the modern textile sector. The increase in low-cost, mass-produced clothing from factories has contributed to wasteful consumerism and regular clothing discard [1, 2]. Fast-fashion retailing is defined by its emphasis on volume sales at discount pricing, encouraging consumers to make frequent clothing purchases. Consumers’ propensity to acquire new items regularly and discard those still wearable promotes a “throwaway culture” [3]. Numerous concerns related to sustainability are brought up by this method used in the textile industry. A strive for resource conservation, and zero waste in the textile industry is essential to achieve sustainability goals. Thus, the textile sector is under pressure to implement sustainable practices to reduce its negative environmental effects.

The increased pressure to implement environmental and sustainability practices has led to considerations for reverse logistics (RL) practices in the textile industry. Following similar RL implementations in sectors such as the electronics industry [4], the automotive industry [5, 6], copier and printer industry [7], RL and recovery practices, such as remanufacturing, refurbishing, or recycling, are gaining prominence in the textile industry [8]. Consequently, recycling firms involved in used clothing collection are putting more effort into activities emphasizing RL via collection and soring, resale, and reuse programs to recover more value from textile waste.

These firms can simultaneously acquire environmental and economic advantages by implementing efficient RL practices. However, the question to be addressed is how these firms can develop RL strategies and continuously improve them [9]. It is impossible to run an RL system efficiently without the ability to measure its performance [10]. Implementing efficient RL requires that firms institute continual measurement of their performance to evaluate different strategies, processes, and capabilities for delivering objectives and developing measures, as well as how they should prioritize the factors that determine their successes [11, 12]. It is important for managing and developing effective practices since it aids and stimulates the decision-making processes involved in RL. RL firms’ performance entails gathering information that leads to smarter choice-making and better strategy development and implementation [9].

The RL literature is still developing [10], and only a few methods have been developed to measure its performance [12 –14]. The most popular method is the balanced scorecard (BSC), which enables managers to examine the firm from different points [13].

However, the balanced scorecard method presents several drawbacks. The method requires a lot of time and dedication to understand, and each evaluation design is tailored to a specific firm. In addition, it is complicated in that it requires a large amount of data, which calls for the participation of many team members. In the course of large participation, data reporting becomes daunting, leading to the presentation of biased and inaccurate figures.

Thus, developing a system to address these drawbacks and improve the operational performances of RL firms is essential. Combining conventional evaluation methods and artificial intelligence (AI) to develop a performance management prediction model helps shorten assessment time, requires the participation of a few team members, and automates the performance evaluation process.

As such, this paper aims to develop a performance prediction model by firstly conducting a conventional benchmarking procedure using data envelopment analysis (DEA). We employ adaptive neuro-fuzzy inference systems (ANFIS) in the second stage to develop a soft computing prediction model. We enhance the capability of ANFIS by different parameters tunning using genetic algorithm (GA), particle swarm optimization (PSO), and Jaya algorithm. The prediction accuracies of each algorithm are compared against the result obtained using DEA to determine their feasibility. To evaluate the model, it is applied in a used clothing collection and recycling firm.

This study is vitally important because it contributes to the RL literature, which is currently in its nascent stages. In particular, it contributes to the limited research on used clothing collection, textile recycling, and RL performance management. Lastly, this study is the first to implement DEA, ANFIS, GA, PSO and Jaya algorithms to assess RL performance.

1.1 Literature review

Compared to forward logistics, the movement of products in RL occurs in the opposite direction [15, 16]. Particularly in textile RL (TRL), the recovery process encompasses the collection, sorting, recycling, and disposal [9 , 18]. Most studies on RL have been conducted on automobiles, scrap metal, e-waste, waste paper, sales packaging, and other materials. However, studies on TRL are not very extensive, and the few available studies concentrate on return handling [9, 19].

The goals of a successful RL enterprise should include lowering costs, implementing performance evaluation to boost output, and making it easier to recycle materials. These aspects are vitally important for firms’ competitiveness [20]. However, calls to implement performance evaluation for RL enterprises have only gained more attention in the supplier selection literature [see [21 –24]]. Some studies have been conducted on RL performance, focusing on recycled computers [25], social commerce [26], the construction industry [27], and other e-waste [20, 28]. However, a literary survey for this paper found no study on the performance evaluation of RL firms focusing on textile waste collection, sorting, and recycling. In addition, we found no study combining data envelopment analysis (DEA), ANFIS, GA, PSO, and Jaya algorithm to develop a soft computing prediction model, as mentioned in section 1.

Multi-criteria decision-making MCDM models such as Analytic Hierarchy Process (AHP) [29, 30], Technique For Order Performance By Similarity To Ideal Solution (TOPSIS) [31, 32], and the Data Envelopment Analysis (DEA) [20 , 34] have been employed in the RL literature. DEA is a technique proposed by Charnes, et al. [35] for assessing the efficiency and performance of a group of homogeneous decision-making units (DMUs) using linear programming. DEA models rank the effectiveness of a group of DMUs based on the utilization of a set of inputs and outputs (outputs) [36]. As a result of its simplicity in implementation, the DEA model has gained widespread acceptance as the most effective benchmarking tool for efficiency assessment. However, its capability is limited—it has no predictive abilities and has to be manually updated each time a criteria changes. Thus, developing a hybrid technique that incorporates DEA and soft computing techniques such as ANFIS enhances the benchmarking capabilities of DEA.

In the Evolutionary Computation (EC) field, PSO and GA are the most popular algorithms used in solving complex optimization problems [37, 38]. They are efficient and possess faster implementation capabilities, which allow them to address problems on a wide scale. They have found widespread usage in the literature on RL [see [39 –41]].

Notwithstanding, limitations abound in using both algorithms—finding the optimum values and solutions for the optimal flow choices in the RL network is difficult. In addition, because both GA and PSO algorithms are based on parameters, their computation time is considerably lengthy, and their results are subject to some biases due to these parameters. To improve the results and shorten computation time, we employ the Jaya algorithm. The Jaya algorithm is a parameterless optimization algorithm recently developed by Venkata Rao [42]. With its lack of parameters, computation times are shorter, and results are bias-free compared to GA and PSO.

2 Methodology

2.1 The DEA model

Data envelopment analysis (DEA) is a widely used analytical technique for assessing performance using resources of homogenous DMUs utilizing a variety of inputs and outputs [36]. DEA identifies if a particular DMU is inefficient or less productive than others. DEA can evaluate the efficiency of any inefficient DMU and provide a benchmark on its efficient frontier.

In this paper, we used the Banker, Charnes and Cooper (BCC) output-oriented model with three output variables and one input variable (see Table 1).

Table 1
Input and output variables for the DEA model

Variable Name Type

Cost (including costs incurred at each center &recycling plant) Input

Quantity of collected textile items Output

Quantity of items processed for reuse Output

Revenue Output

Variable Name	Type
Cost (including costs incurred at each center &recycling plant)	Input
Quantity of collected textile items	Output
Quantity of items processed for reuse	Output
Revenue	Output

Thus, DEA is formulated as: there are n units with s outputs, which is defined by Y_rk, r = 1, . . . s. The input parameter m is defined by X_ik, i = 1 . . . m. Thus, the efficiency measure for DMU_k is defined as follows:

$\begin{matrix} h_{k} = \max \frac{\sum_{r = 1}^{s} u_{r} Y_{rk}}{\sum_{i = 1}^{m} v_{i} X_{ik}} \\ \forall u_{r}, v_{i} ⩾ 0 \end{matrix}$ (1)

To ensure no DMU has an efficiency greater than 1, this constraint is applied to the same weights across all DMUs.

$\frac{\sum_{r = 1}^{s} u_{r} Y_{rj}}{\sum_{i = 1}^{m} v_{i} X_{ij}} ⩽ 1, for j = 1, . . ., n$ (2)

$u_{r} ⩾ 0 for r = 1, . . ., s$ (3)

$v_{i} ⩾ 0 for i = 1, . . ., m$ (4)

$h_{k} = max \sum_{r = 1}^{s} u_{r} Y_{rk}$ (5)

subject to:

$\sum_{i = 1}^{m} v_{i} X_{ij} - \sum_{r = 1}^{s} u_{r} Y_{rj} ⩾ 0 for j = 1, . . . n$ (6)

$\sum_{i = 1}^{m} v_{i} X_{ik} = 1$ (7)

$u_{r} ⩾ 0 for r = 1, . . ., s$ (8)

$v_{i} ⩾ 0 for i = 1, . . ., m$ (9)

2.2 ANFIS

The ANFIS technique has a powerful learning methodology based on a gradient least squares approach used in its reverse spread learning capacity, based on the computation of the derivatives of squared errors, from the output to the input nodes. The least squares approach identifies the forward section’s effect modifiers. The adaptable network consists of five separate layers. The ANFIS structure in Fig. 1 below shows that FIS has two inputs, x and y, and a single output, f, with the levels, nodes, and relationships between them. Thus, the case of the output f and the two inputs, x and y, are as follows [38, 43]:

Fig. 1

The ANFIS structure.

$\begin{matrix} f_{1} = P_{1} + q_{1} y + r_{1}, assume x = A_{1}, y = B_{1} \\ f_{2} = P_{2} + q_{2} y + r_{2}, assume x {= A}_{1}, y = B_{1} \end{matrix}$

All membership grades for each variable are generated by the fuzzification layer, described as:

$O_{1, i} = μ_{Ai} (x) i = 1, 2$ (10)

$O_{1, i} = μ_{Bj} (x) j = 1, 2$ (11)

The fuzzy memberships are denoted by (A_i, B_i), while O_1,i indicates the value produced by the ith node of the first layer (see Equation 12). Nodes of the third layer are calculated using Equation 13, where W_i is determined by the firing strengths of node i, with a normalized firing strength w_i.

$O_{2, i} = W_{i} = μ_{Ai} (x) \times μ_{Bi} (x) i = 1, 2$ (12)

$O_{3, i} = \bar{w} = \frac{W_{i}}{W_{1} + W_{2}}$ (13)

Layer four can be computed in equation 14 below, where P_i, q_i and r_i are consequent parameters:

$O_{4, i} = {\bar{wf}}_{i} = {\bar{w}}_{i} (P_{i} + q_{i} y + r_{i}) i = 1, 2$ (14)

The output at layer five (equation 15) and the final output for the ANFIS (equation 16) can be computed as:

$O_{5, i} = \sum_{i = 1}^{2} {\bar{w}}_{i} f_{i} = \frac{W_{1} f_{1} + W_{2} f_{2}}{W_{1} + W_{2}}$ (15)

$\begin{matrix} Z = & \frac{W_{1}}{W_{1} + W_{2}} f_{1} + \frac{W_{1}}{W_{1} + W_{2}} f_{1} + \dots \\ + \frac{W_{n}}{W_{n - 1} + W_{n}} f_{n} \end{matrix}$ (16)

2.3 PSO

PSO is an example of a population-based search algorithm that mimics the cooperative behavior of a flock of birds. The environment’s fitness information is used to apply various operators to the population, and the result is that the population as a whole should trend toward optimal solutions. Each swarm in the PSO navigates the search area at a speed that is constantly changed based on the accumulated flight (velocity) time of both itself and its companions [44]. The position and velocity of swarms are computed as:

$\begin{matrix} v (k + 1) = v (k) + c_{1} rand (0, 1) \\ \times [pbest (k) - present (k)] + c_{2} rand (0, 1) \\ \times [gbest (k) - present (k) \end{matrix}$ (17)

$present (k + 1) = present (k) + v (k + 1)$ (18)

v(·) denotes the swarm velocity in kth and (k + 1)th iterations. present (·) denotes swarm position. c₁ and c₂ denote learning constants greater than zero, and a random number between [0, 1] is denoted using rand(·) [38].

2.4 Genetic algorithm

In Evolutionary Computation (EC), GA is one of the most popular algorithms for solving complex optimization problems [37, 38]. GA is efficient and possesses faster implementation capabilities solving problems on a wide scale.

Notwithstanding, limitations abound in using PSO and GA algorithms—finding the optimum values and solutions for the optimal flow choices in the reverse logistic network is difficult. In addition, because both GA and PSO algorithms are based on parameters, their computation time is considerably lengthy, and their results are subject to some biases. To improve the results and shorten computation time, we employ the Jaya algorithm.

2.5 Jaya algorithm

Jaya algorithm is a parameter-less optimization algorithm developed by Venkata Rao [42]. With its lack of parameters, computation times are shorter, and results are improved compared to GA and PSO.

Let f (x) be the objective function to be minimized or maximized. Also, ‘n’ number of candidate solutions (k = 1, 2, …, n) be the population size derived from ith iteration with variable size ‘m’ (j = 1, 2, …, m). The best candidate obtains the best value of f (x) (f (x) best), and the worst candidate obtains the worst value of f (x) (f (x) worst). Equation (19) presents a modified value if, during the ith iteration, X_j,k,i is the value of the jth variable for the kth member of a set of potential solutions [45].

$\begin{matrix} X_{j, k, i}^{'} = & X_{j, k, i} + r_{1, j, i} \times (X_{j, best, i} - | X_{j, k, i} |) \\ - r_{2, j, i} \times (X_{j, worst, i} - | X_{j, k, i} |) \end{matrix}$ (19) where X_j,best,i denotes value of the variable j for the best candidate, and X_j,worst,i is the value of the variable j for the worst member of a set of feasible solutions [45].

2.6 Hybrid approach

2.6.1 DEA computation

With these foundations, a hybrid model is developed for evaluating the performance of a TRL firm. As the first step, the efficiencies of DMUs were measured using DEA.

An output orientation BCC model was used in this paper in order to increase outputs of quantities of clothes collected, processed, and the realized revenue, which is the prime objective of a TRL firm. BCC is favored over CCR because, according to Coelli [46], the CRS technology assumption may not always be applicable in certain real-life scenarios and, as a result, cannot be implemented in a wide range of situations.

Thus, the steps involved in the DEA computation of DMUs are as follows:

Identify the input and output variables for measuring the performance of each DMU.

For each DMU, the values of the input and output variables were fetched and computed.

Using equations 1 through 9, a linear programming model for determining the relative performance of each DMU in the RL was performed.

efficiency scores of all DMUs were exported to Matlab version 2021b as input and output data for the ANFIS prediction.

2.6.2 ANFIS prediction

According to Askari and Abbaspour-Gilandeh [47], when input data are susceptible to a somewhat high level of uncertainty, a fuzzy system like ANFIS will be a superior option.

ANFIS is a hybrid intelligent system constructed by connecting a fuzzy logic system with a neural network. Combining fuzzy logic with neural networks in this manner creates a hybrid intelligent system that is effective as it uses the advantages of both fuzzy logic and neural networks.

In the ANFIS method, an input-output data set is used to create a fuzzy inference system (FIS) whose membership function parameters are tweaked using either a back-propagation algorithm or a least squares algorithm. ANFIS learning process involves modifying the membership functions (MFs) and adapting them to a gradient vector. This neuro-adaptive learning process offers a mechanism for fuzzy modeling analogous to neural networks.

With this ANFIS background, the Sugeno system using Gaussian curve membership functions was used in this study.

Thus, in the second step, results obtained from DEA are used as ANFIS inputs and outputs [48]. To enhance ANFIS prediction, three algorithms (GA, Jaya, and PSO) are used for tuning the parameters of ANFIS.

To implement the ANFIS–PSO, a preliminary test based on the coefficient of determination (R²) was performed, changing the numbers of different parameters of the number of particles, the inertia weight, and the cognitive and social learning coefficients (c₁ and c₂). As seen in Table 2, the best performance was obtained from model 4. As a result, both c₁ and c₂ were equated to 2, inertia weight (w) at 1, w_damp at 0.99, and the value of 150 was selected as the number of particles in this study.

Table 2
Parameters of the algorithms used in the study

ANFIS-PSO parameters

Model No. of Particles R ²

Training Testing

1 10 0.964 0.972

2 50 0.950 0.948

3 100 0.974 0.953

4 150 0.981 0.978

5 200 0.932 0.943

ANFIS-GA parameters

Model Population size R ²

Training Testing

1 10 0.971 0.963

2 20 0.981 0.950

3 30 0.978 0.972

4 40 0.972 0.947

5 50 0.946 0.953

ANFIS-PSO parameters
1	10	0.964	0.972
2	50	0.950	0.948
3	100	0.974	0.953
4	150	0.981	0.978
5	200	0.932	0.943
ANFIS-GA parameters
Model	Population size	R ²
		Training	Testing
1	10	0.971	0.963
2	20	0.981	0.950
3	30	0.978	0.972
4	40	0.972	0.947
5	50	0.946	0.953

A similar test was performed for the ANFSI-GA implementation to determine the ideal mutation rate, mutation percentage (pm), crossover percentage (pc), selection pressure, and number (size) of the population. Model 3, with 30 population size, was selected based on R² values (see Table 2). In addition, the crossover percentage was given 0.4, mutation rate 0.15, mutation percentage 0.7, and selection pressure 8.

Furthermore, the ANFIS-PSO, ANFIS-Jaya, and ANFIS-GA had maximum iterations of 1000.

2.7 Case study

Data for the model were gathered from a textile recovery firm. The company operates a large network of textile collection centers across several cities. Materials collected are sorted into several categories. The reusable clothing items are graded into three classes (A, B, and C) and exported to other countries as second-hand clothes (SHC). Those falling short of this classification are upcycled, remanufactured into carpets, wiping cloth, or landfilled.

Data for such a reverse logistic operation is difficult to collect and maintain. In addition, continuity in data collection is difficult as some collection centers are moved due to construction or population shift. However, data covering a period of three years across four cities, each with 18 collection centers, were available. Each of the centers, the cities and the years were taken as a DMU, yielding 216 total DMUs, using key indicators from its RL network (see Table 1).

3 Results

There were 216 data points for use as input and output. To prevent overfitting, the data were partitioned into 173 data for training and 43 for testing.

The accuracies of the prediction models were evaluated using several standard performance criteria. The root mean square (RMSE), the mean absolute percentage error (MAPE), the coefficient of determination (R²), and the mean absolute error (MAE) (equations 20 –23) were used to evaluate the effectiveness of the models.

$RMSE = \sqrt{\frac{\sum_{i = 1}^{n} {({\hat{y}}_{i} - y_{i})}^{2}}{n}}$ (20)

$MAPE = \frac{1}{n} \sum_{i = 1}^{n} | \frac{{\hat{y}}_{i} - y_{i}}{y_{i}} | \times 100$ (21)

$R^{2} = 1 - \frac{\sum_{i = 1}^{n} {(y_{i} - {\hat{y}}_{i})}^{2}}{\sum_{i = 1}^{n} {(y_{i} - \overset{ˇ}{y})}^{2}}$ (22)

$MAE = \frac{1}{n} \sum_{i = 1}^{n} | {\hat{y}}_{i} - y_{i} |$ (23) where n denotes the number of data sets, $\hat{y}$ represents the predicted values, and y represents the actual values obtained from DEA.

As seen in Table 3, all the algorithms have improved accuracies compared to the ANFIS. However, each algorithm presented minor differences in their performance measures in predicting the reverse logistics performance. During the training and testing of the models, taking into consideration the RMSE, MAE, and MAPE indices, as well as R², revealed that the ANFIS–Jaya model performed better than the others in predicting the performance of this reverse logistics. The ANFIS–Jaya model was shown to be superior to both the ANFIS–GA, and ANFIS–PSO models, even though all of these models were found to be adequately capable of predicting the performance of reverse logistics. In addition, the ANFIS–Jaya model produced the lowest RMSE, MAE, and MAPE results, while ANFIS-GA presented better results than ANFIS-PSO.

Table 3

Examining the similarities and differences in the performance of prediction models in the training and testing phases

Models	RMSE		MAPE		MAE
	Training	Testing	Training	Testing	Training	Testing
ANFIS	0.062	0.059	9.451	9.544	0.510	0.498
ANFIS-PSO	0.050	0.054	7.395	9.877	0.332	0.491
ANFIS-GA	0.032	0.038	6.660	5.384	0.201	0.232
ANFIS-Jaya	0.030	0.036	2.981	3.091	0.157	0.163

The superior performance of ANFIS-Jaya can be interpreted by its flexibility and parameterless characteristics, which empower the algorithm to converge to global optima in a shorter time.

To compare the coefficient of determination (R²), a scatter plot, a simple statistical tool for depicting the interrelationship of several variables, was used. It is often combined with a simple linear regression line to fit a model between the two parameters.

The resulting cross-fit line in Fig. 2 shows the ANFIS model’s accuracies with R² values of 0.8885 in the training and 0.8966 in the testing and datasets.

Fig. 2

Scatterplot of predicted vs actual efficiency scores from ANFIS and hybrid PSO models for both training and testing.

With these scores, a general conclusion cannot be made that the ANFIS prediction was much better than that of the DEA. However, introducing the particle swarm optimization algorithm was able to tweak the parameters of ANFIS to enhance its accuracy. As seen in Fig. 2, the ANFIS-PSO hybrid model produced a much-improved accuracy with R² values of 0.9516 and 0.9554 in training and testing datasets, respectively.

In addition, these values were compared against the Jaya and Genetic algorithms (Fig. 3). Using GA to tweak the ANFIS model produced a better prediction accuracy than both ANFIS and PSO with R² values of 0.9690 and 0.9712 in training and testing, respectively.

Fig. 3

Scatterplot of predicted vs actual efficiency scores from ANFIS hybrid models of GA and Jaya algorithms for both training and testing.

As mentioned in section 2.5, the Jaya algorithm is a parameterless optimization algorithm developed by Venkata Rao [42]. With its lack of parameters, computation times are shorter, and results are improved compared to GA and PSO.

The result of the ANFIS-Jaya prediction model revealed the highest prediction improvement. The ANFIS-Jaya hybrid model generated R² values of 0.9832 and 0.9852 in training and testing datasets, respectively (see Fig. 3).

4 Discussion

To begin with model construction, we first used DEA to determine the levels of DMU efficiency. Data covering three years across four cities, each with 18 collection centers, and each of the centers, the cities, and the years were taken as a DMU, yielding 216 total DMUs, using key indicators from the RL network. An output orientation BCC model was used in this paper in order to increase outputs of quantities of clothes collected, processed, and the realized revenue. We further introduced a fuzzy system to enhance the efficiencies of DMUs by introducing ANFIS. ANFIS is a kind of adaptive multi-layer feedforward network that uses historical data to make predictions about future data, using a combination of neural network and fuzzy inference system capabilities. Thus, in our ANFIS system, fuzzy rules are automatically extracted by the neural networks from efficiency scores generated using the DEA model. During the training process, the membership functions undergo adaptive adjustment. This makes it easier to predict the ranking of new DMUs based on the ranking and efficiency of previous DMUs without performing many calculations as is required by the DEA method.

This paper employed PSO, GA, and Jaya algorithms to adjust the ANFIS parameters and membership functions. From the results and analysis of the previous section, it can be observed that PSO and GA obtained good accuracy rates. However, in comparison, the Jaya algorithm produced matrices whose mean accuracies were the highest among all tested models. The reported accuracy rates of our hybrid models were superior to those previously published, such as in [48, 49]. However, our computing times were relatively longer than reported in [48, 49]. Emphatically, the PSO model was found to be more time-consuming than other models. This can be largely attributed to differences in computer architecture and the use of other programming environments different from ours. Other computational time reduction methods presented in [50] and the use of parallel technologies for the reduction of processing time and costs can be used to avoid these challenges. Moreover, the Jaya algorithm model had the advantage of being less computationally expensive for any given size of network topology during our experiment, particularly for triangular membership functions.

5 Conclusion

Implementing reverse logistics processes in the textile industry is one of the least studied areas. However, due to the rising need to embrace environmentally responsible and sustainable business practices, RL and recovery practices, such as remanufacturing, refurbishing, or recycling, are beginning to gain prominence in the textile industry. These practices follow in the footsteps of similar RL implementations in the electronics, automotive, copier, and printer industries. Recycling companies engaged in collecting used clothing are devoting more resources to operations emphasizing RL. These activities include collection and sorting programs and resale and reuse initiatives.

Using these processes, textile recycling companies can simultaneously obtain economic and environmental benefits via more efficient RL practices. However, a system for measuring these efficiencies is paramount, as it is impossible to run an RL system efficiently without the ability to measure its performance. Thus, implementing efficient RL requires that firms institute continual measurement of their performance to evaluate different strategies, processes, and capabilities for delivering objectives and developing measures, as well as how they should prioritize the factors that determine their success.

There are many performance assessment methods; however, they are associated with many limitations.

Understanding their process takes time and effort, and each assessment is designed to meet a particular entity. The balanced scorecard, in particular, is difficult to implement because it necessitates collecting a substantial quantity of data, which demands the involvement of many workers. The reporting of data becomes challenging when there is a great amount of involvement, which may lead to the display of results that are skewed and erroneous.

Thus, developing a strategy to overcome these shortcomings and enhance RL processes is crucial. One effective method is by combining traditional evaluation techniques with AI to construct a performance prediction model to shorten assessment time and automate the performance review.

In this study, a model for performance prediction has been developed to aid the efficiency assessment of textile reverse logistics firms. Data from a TRL firm covering three years across four cities, each with 18 collection centers, were used. The DEA analysis was conducted by taking each of the centers, the cities, and the years as a DMU. This process yielded 216 total DMUs, using key RL indicators. Results from the DEA computation were used as input and output for the ANFIS prediction. To enhance the ANFIS prediction, PSO, GA, and the Jaya algorithms were introduced. Results from the four models revealed that ANFIS-Jaya has a better prediction accuracy with R² values of 0.9832 and 0.9851 in training and testing datasets, respectively.

This study is vitally important because it contributes to the RL literature. In particular, it contributes to the limited research on used clothing collection, textile recycling, and RL performance management. In addition, this study is one of the first to use the DEA, ANFIS, GA, PSO, and Jaya algorithms to evaluate RL’s efficiencies.

References

Gazzola

, Pavione

, Pezzetti

and Grechi

, Trends in the Fashion Industry. The Perception of Sustainability and Circular Economy: A Gender/Generation Quantitative Approach, Sustainability 12(7) (2020), 2809. [Online]. Available: https://www-mdpi-com-s.web.bisu.edu.cn/2071-1050/12/7/2809

Hawley

J.M.

, Understanding and improving textile recycling: a systems perspective, Elsevier, (2009), pp. 179–199.

Joung

H.-M.

, Fast-fashion consumers’ post-purchase behaviours, International Journal of Retail & Distribution Management 42 (2014), 688–697. doi: 10.1108/IJRDM-03-2013-0055

Atasu

and Subramanian

, Extended Producer Responsibility for E-Waste: Individual or Collective Producer Responsibility? Production and Operations Management 21(6) (2012), 1042–1059. doi: https://doi.org/10.1111/j.1937-5956.2012.01327.x

Kaviani

M.A.

, Tavana

, Kumar

, Michnik

, Niknam

and d. Campos

E.A.R.

, An integrated framework for evaluating the barriers to successful implementation of reverse logistics in the automotive industry, Journal of Cleaner Production 272 (2020), 122714. doi: https://doi.org/10.1016/j.jclepro.2020.122714

Makarova

, Shubenkova

and Pashkevich

, The Concept of the Decision Support System to Plan the Reverse Logistics in Automotive Industry, in 2018 26th International Conference on Software, Telecommunications and Computer Networks (SoftCOM), 13–15 Sept. 2018, 2018, pp. 1–6. doi: 10.23919/SOFTCOM.2018.8555760

Savaskan

R.C.

and Wassenhove

L.N.V.

, Reverse Channel Design: The Case of Competing Retailers, Management Science 52(1) (2006), 1–14. doi: 10.1287/mnsc.1050.0454

Pal

, Value creation through reverse logistics in used clothing networks, The International Journal of Logistics Management 28(3) (2017), 864–906. doi: 10.1108/IJLM-11-2016-0272

Ahlström

, Ferning

, Cheniere

M.K.

and Sorooshian

, Performance Indicators of Textile Reverse Logistics, (in en), IOP Conf Ser: Earth Environ Sci 442(1) (2020), 012012. doi: 10.1088/1755-1315/442/1/012012

10.

Butzer

, Schötz

, Petroschke

and Steinhilper

, Development of a Performance Measurement System for International Reverse Supply Chains, Procedia CIRP 61 (2017), 251–256. doi: https://doi.org/10.1016/j.procir.2016.11.264

11.

Genchev

S.E.

, Glenn Richey

and Gabler

C.B.

, Evaluating reverse logistics programs: a suggested process formalization, The International Journal of Logistics Management 22(2) (2011), 242–263. doi: 10.1108/09574091111156578

12.

Shaik

M.N.

and Abdul-Kader

, Comprehensive performance measurement and causal-effect decision making model for reverse logistics enterprise, Computers & Industrial Engineering 68 (2014), 87–103. doi: https://doi.org/10.1016/j.cie.2013.12.008

13.

Shaik

and Abdul-Kader

, A comprehensive performance measurement framework for reverse logistics enterprise, 61st Annual IIE Conference and Expo Proceedings, 01/01 2011.

14.

Janse

, Schuur

and de Brito

M.P.

, A reverse logistics diagnostic tool: the case of the consumer electronics industry, The International Journal of Advanced Manufacturing Technology 47(5) (2010), 495–513. doi: 10.1007/s00170-009-2333-z

15.

Tibben-Lembke

R.S.

and Rogers

D.S.

, Differences between forward and reverse logistics in a retail environment, Supply Chain Management: An International Journal 7(5) (2002), 271–282. doi: 10.1108/13598540210447719

16.

Zarbakhshnia

, Soleimani

, Goh

and Razavi

S.S.

, A novel multi-objective model for green forward and reverse logistics network design, Journal of Cleaner Production 208 (2019), 1304–1316. doi: https://doi.org/10.1016/j.jclepro.2018.10.138.

17.

Daaboul

, Le Duigou

, Penciuc

and Eynard

, Reverse logistics network design: a holistic life cycle approach, Journal of Remanufacturing 4(1) (2014), 7. doi: 10.1186/s13243-014-0007-y

18.

Cullinane

, Karlsson

, Browne

and Yingli

, Retail clothing returns: A review of key issues, 2017, 2017, doi: 10.13140/RG.2.2.35163.46882

19.

Jain

, A three-stage hybrid integrated decision making framework for modelling reverse logistics operations: a case of a textile company, (in en), IJISE 23(3) (2016), 370. doi: 10.1504/IJISE.2016.076869

20.

Wang

, Hao

, Gao

, Zhang

and Zhou

, Multi-attribute decision making on reverse logistics based on DEA-TOPSIS: A study of the Shanghai End-of-life vehicles industry, Journal of Cleaner Production 214 (2019), 730–737. doi: https://doi.org/10.1016/j.jclepro.2018.12.329

21.

Meade

and Sarkis

, A conceptual model for selecting and evaluating third-party reverse logistics providers, Supply Chain Management: An International Journal 7(5) (2002), 283–295. doi: 10.1108/13598540210447728

22.

Keshavarz Ghorabaee

, Amiri

, Olfat

and Khatami Firouzabadi

S.M.A.

, Designing a multi-product multi-period supply chain network with reverse logistics and multiple objectives under uncertainty, Technological and Economic Development of Economy 23(3) (2017), 520–548. doi: 10.3846/20294913.2017.1312630

23.

Tavana

, Shaabani

, Santos-Arteaga

F.J.

and Valaei

, An integrated fuzzy sustainable supplier evaluation and selection framework for green supply chains in reverse logistics, Environmental Science and Pollution Research 28(38) (2021), 53953–53982. doi: 10.1007/s11356-021-14302-w.

24.

Zhang

, Li

and Wang

, A Review of the Criteria and Methods of Reverse Logistics Supplier Selection, Processes 8(6) (2020), 705. [Online]. Available: https://www-mdpi-com-s.web.bisu.edu.cn/2227-9717/8/6/705

25.

Huang

R.-H.

, Yang

C.-L.

, Lin

C.-C.

and Cheng

Y.-J.

, Performance evaluation model for reverse logistics-the case of recycled computers, Journal of Statistics and Management Systems 15(2-3) (2012), 323–343. doi: 10.1080/09720510.2012.10701629

26.

Han

and Trimi

, A fuzzy TOPSIS method for performance evaluation of reverse logistics in social commerce platforms, Expert Systems with Applications 103 (2018), 133–145. doi: https://doi.org/10.1016/j.eswa.2018.03.003

27.

Hammes

, De Souza

E.D.

, Taboada Rodriguez

C.M.

, Rojas Millan

R.H.

and Mojica Herazo

J.C.

, Evaluation of the reverse logistics performance in civil construction, Journal of Cleaner Production 248 (2020), 119212. doi: https://doi.org/10.1016/j.jclepro.2019.119212

28.

Bansia

, Varkey

J.K.

and Agrawal

, Development of a Reverse Logistics Performance Measurement System for a Battery Manufacturer, Procedia Materials Science 6 (2014), 1419–1427. doi: https://doi.org/10.1016/j.mspro.2014.07.121

29.

Bouzon

, Govindan

, Rodriguez

C.M.T.

and Campos

L.M.S.

, Identification and analysis of reverse logistics barriers using fuzzy Delphi method and AHP, Resources, Conservation and Recycling 108 (2016), 182–197. doi: https://doi.org/10.1016/j.resconrec.2015.05.021

30.

Tavana

, Zareinejad

, Di Caprio

and Kaviani

M.A.

, An integrated intuitionistic fuzzy AHP and SWOT method for outsourcing reverse logistics, Applied Soft Computing 40 (2016), 544–557. doi: https://doi.org/10.1016/j.asoc.2015.12.005

31.

Sirisawat

and Kiatcharoenpol

, Fuzzy AHP-TOPSIS approaches to prioritizing solutions for reverse logistics barriers, Computers & Industrial Engineering 117 (2018), 303–318. doi: https://doi.org/10.1016/j.cie.2018.01.015

32.

Prakash

and Barua

M.K.

, Integration of AHP-TOPSIS method for prioritizing the solutions of reverse logistics adoption to overcome its barriers under fuzzy environment, Journal of Manufacturing Systems 37 (2015), 599–615. doi: https://doi.org/10.1016/j.jmsy.2015.03.001

33.

Song

, Ma

and Chen

, A Profit Distribution Model of Reverse Logistics Based on Fuzzy DEA Efficiency—Modified Shapley Value, Sustainability 13(13) (2021), 7354. [Online]. Available: https://www-mdpi-com-s.web.bisu.edu.cn/2071-1050/13/13/7354

34.

Momeni

, Tavana

, Mirzagoltabar

and Mirhedayatian

S.M.

, A new fuzzy network slacks-based DEA model for evaluating performance of supply chains with reverse logistics, Journal of Intelligent & Fuzzy Systems 27 (2014), 793–804. doi: 10.3233/IFS-131037

35.

Charnes

, Cooper

W.W.

and Rhodes

, Measuring the efficiency of decision making units, European Journal of Operational Research 2(6) (1978), 429–444. doi: https://doi.org/10.1016/0377-2217(78)90138-8

36.

Zahedi-Seresht

, Khosravi

, Jablonsky

and Zykova

, A data envelopment analysis model for performance evaluation and ranking of DMUs with alternative scenarios, Computers & Industrial Engineering 152 (2021), 107002. doi: https://doi.org/10.1016/j.cie.2020.107002

37.

Alshamsi

and Diabat

, A Genetic Algorithm for Reverse Logistics network design: A case study from the GCC, Journal of Cleaner Production 151 (2017), 652–669. doi: https://doi.org/10.1016/j.jclepro.2017.02.096

38.

Shamshirband

, Hadipoor

, Baghban

, Mosavi

, Bukor

and Várkonyi-Kóczy

A.R.

, Developing an ANFIS-PSO Model to Predict Mercury Emissions in Combustion Flue Gases, Mathematics 7(10) (2019), 965. [Online].Available: https://www-mdpi-com-s.web.bisu.edu.cn/2227-7390/7/10/965

39.

, Li

and Liang

, Multi-objective optimization of reverse logistics network based on improved particle swarm optimization, in 2008 7th World Congress on Intelligent Control and Automation, 25–27 June 2008, 2008, pp. 7476–7480. doi: 10.1109/WCICA.2008.4594085

40.

Rachih

, Mhada

F.Z.

and Chiheb

, Meta-heuristics for reverse logistics: A literature review and perspectives, Computers & Industrial Engineering 127 (2019), 45–62. doi: https://doi.org/10.1016/j.cie.2018.11.058

41.

Guo

, Wang

, Fan

and Gen

, Forward and reverse logistics network and route planning under the environment of low-carbon emissions: A case study of Shanghai fresh food E-commerce enterprises, Computers & Industrial Engineering 106 (2017), 351–360. doi: https://doi.org/10.1016/j.cie.2017.02.002

42.

Venkata Rao

, Jaya: A simple and new optimization algorithm for solving constrained and unconstrained optimization problems, 10.5267/j.ijiec, pp. 19–34, 2016. doi: 10.5267/j.ijiec.2015.8.004

43.

Qin

, Langari

and Gu

, A new modeling algorithm based on ANFIS and GMDH, Journal of Intelligent & Fuzzy Systems 29 (2015), 1321–1329. doi: 10.3233/IFS-141443

44.

Kennedy

and Eberhart

, Particle swarm optimization, in Proceedings of ICNN’95 –International Conference on Neural Networks, 27 Nov.-1 Dec. 1995 4 (1995), 1942–1948. doi: 10.1109/ICNN.1995.488968

45.

Banerjee

and Sarkar

, Modified Jaya Optimization Algorithm for Combined Economic Emission Dispatch Solution, International Journal of Electrical Energy 6 (2018), 13–19. doi: 10.18178/ijoee.6.1.13-19

46.

Coelli

, A multi-stage methodology for the solution of orientated DEA models, Operations Research Letters 23(3) (1998), 143–149. doi: https://doi.org/10.1016/S0167-6377(98)00036-4

47.

Askari

and Abbaspour-Gilandeh

, Assessment of adaptive neuro-fuzzy inference system and response surface methodology approaches in draft force prediction of subsoiling tines, Soil and Tillage Research 194 (2019), 104338. doi: https://doi.org/10.1016/j.still.2019.104338

48.

Khaldi

, Afia

A.E.

and Chiheb

, Performance prediction of pharmaceutical suppliers: comparative study between DEA-ANFIS-PSO and DEA-ANFIS-GA, International Journal of Computer Applications in Technology 60(4) (2019), 317–325. doi: https://doi.org/10.1504/IJCAT.2019.101172

49.

Pousinho

H.M.I.

, Mendes

V.M.F.

and Catalão

J.P.S.

, A hybrid PSO–ANFIS approach for short-term wind power prediction in Portugal, Energy Conversion and Management 52(1) (2011), 397–402. doi: https://doi.org/10.1016/j.enconman.2010.07.015

50.

Parra

, Galan

S.G.

, Yuste

A.J.

, Prado

R.P.

and Muñoz

J.E.

, Method to Minimize Distributed PSO Algorithm Execution Time in Grid Computer Environment, Berlin, Heidelberg, 2009: Springer Berlin Heidelberg, in Bioinspired Applications in Artificial and Natural Computation, pp. 478–487. doi: https://doi.org/10.1007/978-3-642-02267-8_51

ANFIS-PSO parameters
Model	No. of Particles	R ²
		Training	Testing
1	10	0.964	0.972
2	50	0.950	0.948
3	100	0.974	0.953
4	150	0.981	0.978
5	200	0.932	0.943
ANFIS-GA parameters
Model	Population size	R ²
		Training	Testing
1	10	0.971	0.963
2	20	0.981	0.950
3	30	0.978	0.972
4	40	0.972	0.947
5	50	0.946	0.953