Decision Analytics Journal 7 (2023) 100227

Contents lists available at ScienceDirect

Decision Analytics Journal

journal homepage: www.elsevier.com/locate/dajour

A multi-agent coalition-based approach for order fulfilment in e-commerce

Pragyan Nanda a ,∗, Srikanta Patnaik b
a
SOA University, Bhubaneswar, Odisha, India
b
Interscience Institute of Management and Technology, Bhubaneswar, Odisha, India

ARTICLE INFO ABSTRACT

Keywords: The rapid growth of e-commerce in the recent pandemic is driven by the wide adoption of digitization,
E-commerce marketplace extensive use of smartphones, affordable internet services, and the convenience of shopping from anywhere and
Order fulfilment anytime. Some key indexes for measuring the performance of the e-commerce organization include affordable
Coalition-formation
deals, competitive prices, and efficiency in fulfilling orders while ensuring customer satisfaction. Customer
Cost minimization
review ratings are one of the significant metrics to measure customer satisfaction, influencing other customers
Review score
Lead time
to make purchase decisions. Therefore, favourable customer reviews and ratings of various products presented
on e-commerce platforms provide competitive advantages to sellers. Usually, customers placing an order on
an e-commerce platform are notified about the estimated delivery date. Consequently, after receiving the
product delivery, customers give feedback in the form of review comments and ratings to communicate their
satisfaction or dissatisfaction towards the seller. In case of delays in order fulfilment, the customers mostly give
negative feedback along with poor ratings, which can affect the seller’s reputation, thus impacting future sales.
Therefore, the central idea of this research is to adopt a resilient order fulfilment approach by identifying the
optimal set of sellers based on fulfilment cost, lead time for fulfilment and review rating score obtained from
previous orders. The research work has been divided into three different phases. Firstly, significant constraints
associated with an e-commerce fulfilment system have been identified from existing literature. The relationship
between the lead time in order fulfilment and review rating score is studied. Finally, a multi-agent coalition-
based framework is proposed for generating the Pareto-optimal set of reliable sellers to fulfil an order. Although
sellers in an e-commerce marketplace are usually considered competitors of each other, a temporary coalition
among the sellers for aggregated order fulfilment can minimize fulfilment costs. The proposed multi-agent
coalition-based framework is implemented on real-time data to confirm the findings.

1. Introduction and interactions. This high-level interdependency may exist among

various products and their markets, design and development processes,
The e-commerce business has recently become the most cutting- management practices, and the business ecosystem consisting of the
edge and popular one in the entire world. As a result of the COVID-19 functional aspects. Irrespective of the fact that business organizations
epidemic, the e-commerce store market on a global scale has become deal with essential or non-essential products or even services, several
significantly more active than the conventional in-commerce industry. external and internal factors are responsible for making the current
People’s increased internet use during this period has encouraged them
business environment highly complex. Such factors as globalization,
to do more shopping online, where they can avoid wasting time and
availability of granular information, exploration of new product mar-
energy. Again, with globalization, growing digitalization, availability
kets, volatility due to fragmented markets, internal management and
of granular data and advancements in information technology, the
administration of various interdependent and overlapping processes
business environment is becoming highly complicated at an exponential
rate [1–3]. This growing complicacy makes it increasingly difficult to and their layered relationships with each other increase complexity
analyse, plan and manage various core processes of the organizations. with an increase in the number of customers [6–8]. Some common
Also, most traditional management tools fail to capture this emerging factors influencing the complexity include the range of products of-
complexity due to a lack of robustness. One highly complex, dynamic fered, the number of operational processes involved, the number of
and unpredictable sector is the e-commerce sector which has been operational business locations, the nature of the operational work being
seeing exponential growth while putting forward new opportunities carried out at each location and the number of people involved in the
and challenges [4,5]. Complexity in an e-commerce business organi- operational processes [9,10]. Furthermore, the continuously changing
zation is a by-product of many highly interdependent relationships behaviour of customers, volatility and unpredictability of the consumer

∗ Corresponding author.
E-mail addresses: [email protected] (P. Nanda), [email protected] (S. Patnaik).

https://doi.org/10.1016/j.dajour.2023.100227
Received 14 November 2022; Received in revised form 7 April 2023; Accepted 7 April 2023
Available online 18 April 2023
2772-6622/© 2023 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
P. Nanda and S. Patnaik Decision Analytics Journal 7 (2023) 100227

market due to uncertainty and pressure of cut-throat competition also time. Therefore, solutions, including trade-offs for meeting competitive
fuel the complexity in modern business organizations. objectives, are the goal of optimizing the order fulfilment process [33–
Mainly in the e-commerce market, uncertainty highly impacts con- 35]. However, in order to minimize the transportation cost, coalitions
sumers’ purchase decisions; therefore, sellers attempt to build a positive are formed among sellers so that the orders from a single customer
relationship with the customers so that they can return for future can be merged at one location confined within a collaborative setting
purchases. Furthermore, the fluctuation in customer demands, which for fulfilment [36–39]. The overall solution of the proposed approach
contributes a lot to increased uncertainty, is driven by many factors focuses on a multi-agent-based coalition formation framework inspired
such as the quality of the end product, comparative price of the by the crocodile predatory strategy presented by Nanda et al. [40].
products, and average lead time that they have to wait after placing The proposed framework forms a coalition among sellers with mutually
the order to receive the product [11–14]. Customer review scores are agreed upon conditions in terms of rank order based on the average
one of the most visible metrics to measure customer satisfaction to lead time taken by each seller to fulfil the orders and review score
anticipate their loyalty towards specific sellers or brands. A higher obtained for each order, along with the sharing of the delivery cost
review score indicates the customer’s trust and positive satisfaction for a win–win situation.
level ([6,15,16]). In contrast, lower scores can indicate dissatisfaction Thus, the proposed approach presents the order fulfilment problem
due to multiple reasons ranging from delay in order delivery to the as a multi-objective optimization problem where the optimization goal
lousy quality of the product or even missing products [17–20]. Also, includes three different and conflicting objectives, such as the delivery
today’s shoppers anticipate finding extensive product selections in the cost needs to be minimized and the average lead time needs to be
e-commerce markets they frequent. minimized. In contrast, the review score needs to be maximized. In
Further, they anticipate receiving goods acquired online within a such a scenario, no single solution can be considered an ideal solution,
couple of days as soon as possible at the place of their choice. Therefore, but instead, a set of optimal solutions can be generated according to
to maintain a competitive advantage in an on-demand market, a given the Pareto front Method. Generally, Pareto optimum solutions present
organization needs to have a distributed order fulfilment system that is a set of non-dominated solutions with trade-offs between distinct and
both efficient and flexible. A business’s process from receiving an order conflicting objectives [41–43]. In other words, no solution exists that
until it is delivered to the customer is known as the order fulfilment improvises one objective without deteriorating another [44,45]. There-
process [21–24]. It involves the sub-process of collecting goods from fore, the proposed coalition framework focuses on generating Pareto
the warehouses, then processing them further by packaging them and optimum solutions for order fulfilment purposes while minimizing
finally delivering them to the customers. It can be expensive and the fulfilment cost and the average lead time and maximizing the
complex if not followed by optimal strategies. Ensuring that the correct customer review. The framework has been implemented on real-time
product has been delivered to the correct customer in a time-efficient data to demonstrate its effectiveness in generating Pareto optimal solu-
manner is quite a challenge in large-scale e-commerce environments tions, and the outcomes have been discussed. The proposed framework
[25]. In large-scale settings, where hundreds of thousands of orders are outperforms the existing fulfilment approach available in the litera-
placed by customers daily, the order fulfilment process has numerous ture. The overall performance can be improvised in terms of reduced
aspects of operations to be considered. It includes gathering orders, fulfilment cost, increased revenue and reputation.
locating those goods in warehouses and distribution centres, and pack- Additionally, Multi-agent Systems (MAS), comprised of several in-
aging and assignment of shipments to correct delivery options [26]. telligent agents that interact with one another, have emerged as a new
However, Order fulfilment options may directly influence a customer’s paradigm for the modelling, creating, and implementing of software
buying behaviour. Most Customers want to have delivery options to solutions for the tackling of complex and dispersed problems [33,42].
suit their requirements. In order to gain competitive advantages, or- Both the implementation of multi-agent systems and their use have
ganizations usually offer a combination of fast and moderate shipping been significant contributors to academic study. The proliferation of the
options to satisfy consumers’ budgets and time constraints. Therefore, internet has led to an increase in the number of people using apps based
order fulfilment strategy can be approached from various angles for a on agents. A multi-agent system boosts overall system performance.
targeted group of customers with their advantages. It offers dependability, extensibility, robustness, maintainability, re-
Again, in the context of e-commerce business, products must be sponsiveness, adaptability, reuse, and flexibility, among other benefits.
transhipped between warehouses or transported to distribution centres In a multi-agent system, the agents are highly developed computer
located at different locations to fulfil the orders cost-effectively [27]. programmes capable of acting independently on behalf of their users
E-commerce firms adopt different strategies to manage the volatility across various settings. It can communicate either with centralized
in customer demands by arranging inventories appropriately while agents or among themselves, depending on the Design. Software agents
optimizing the order fulfilment process to minimize expenses and main- have the potential to revolutionize trading in electronic commerce by
tain high service levels [28]. Proper design and administration of the assisting buyers in making purchases that are tailored to their own
order fulfilment process can contribute to achieving a higher per- needs, interests, and preferences. E-commerce platforms are realizing
formance for e-commerce firms [29]. The order fulfilment process that it is critical to provide individualized shopping experiences for
involves optimization of the coordination among people, processes customers to bring additional value to their products and services.
and sub-processes, resources and information shared among them for In our opinion, the next step in developing e-commerce platforms is
successful accomplishment. It involves the manufacturing plants, sup- to emphasize supporting collaborative efforts [34,37,46]. The primary
pliers, sellers, warehouses, distribution centres and final markets or objective of this project is to develop a collaborative multi-agent-based
customers [30]. E-commerce order fulfilment processes are more chal- e-commerce framework. This framework should allow for autonomy,
lenging and complex than traditional fulfilment systems due to the pro-activity, and personalization and include an intelligent agent that
wide distribution of the customer base. Therefore, optimizing the order can monitor each other. The purpose of the suggested framework is to
fulfilment process makes it faster and more resilient by incorporating give meaningful responses to meaningful requests and to offer relevant
flexibility while maintaining the fulfilment cost as low as possible [31, goods to individuals who need it when they need it in a manner that
32]. In order to model a resilient order fulfilment system, multiple ob- satisfies their interests in a way that takes into account their prefer-
jectives need to be achieved together, such as minimizing the lead time, ences. In order to illustrate the viability of the proposed framework,
minimizing the fulfilment cost and maximizing the inventory level to a prototype will be developed and evaluated. The remaining content
avoid stock-out delays. Since, in some occasions, these objectives may has been organized into the following sections as Section 2 discusses
conflict with each other such as maximizing inventory stock may incur some recent related works. Section 3 discusses various challenges en-
increased cost while reducing inventory may result in increased lead countered in the order fulfilment process considered from the existing

2
P. Nanda and S. Patnaik Decision Analytics Journal 7 (2023) 100227

literature; Section 4 identifies various constraints and formulates the Again, while e-commerce has revolutionized the retail and logistics
mathematical modelling of the constraints. Next, Section 5 discusses industry, customers are empowered to place orders from any corner of
the Proposed Multi-Agent Framework is presented in detail, Section 6 the world. Despite the enormous positive growth of the e-commerce
The real-time dataset considered has been discussed in this Section 7 industry, logistic service providers need help complying with the in-
in this section results and findings are discussed; and lastly, Section 8 creasing and diversified demands of the customers. Order fulfilment
presents the conclusion and future work. is an area where e-commerce organizations need more bottlenecks.
In addition to previously discussed literature, Zhao et al. study the
2. Literature review order fulfilment cost minimization problem in a multi-item order ful-
filment scenario where the e-retailer has the responsibility to decide
E-commerce organizations, while emerging as a new trend in shop- on the selection of the warehouse(s) to fulfil each order in the pres-
ping, provide the circulation of several commodities such as groceries, ence of various relevant constraints [49]. Nevertheless, the boom of
apparel, utilities, gifts, cosmetics and so on among various entities such e-commerce has resulted in e-retailers expanding fulfilment centres
as retail sellers, wholesalers, manufacturers and customers. They not for faster deliveries and significantly increasing the fulfilment cost in
only bring billions of transactions across every corner of the globe contrast. Their research mainly focuses on a two-layered distribution
but also project significant bottlenecks in order fulfilment that affect network containing regional distribution centres (RDCs) that are larger
the global e-business and customer e-shopping experience. It involves in the upper layer and front distribution centres (FDCs) that are smaller
accepting the orders placed by the customers and effectively fulfilling in size in the lower layer. In such an experimental set-up, they analyse
the orders by timely delivery of the products requested in the placed the performance of myopic policies for choosing the minimal cost
orders. Therefore, order fulfilment can be considered the backbone of fulfilment option for each order. They further consider an optimal
any e-commerce organization. Also, customer satisfaction depends on clairvoyant policy to compare their approach and empirically estimate
the timely delivery of products to some extent which may otherwise the upper bound of the performance ratio using FedEx shipping rates.
lead to the cancellation of orders with a deterioration of reputation They then extend their study to prove the asymptotic optimality of
as a penalty. E-commerce fulfilment policies have drawn a significant linear program rounding policy in a demand forecasting set-up and
amount of attention from researchers with the increase in the growth complement the results with numerical experimentation.
of online shopping. For more minor organizations, order fulfilment Again, in their work, Bansal et al. have studied the performance
is relatively easy. It can be easily accomplished, but for large-scale analysis of batching decisions in order fulfilment fluctuations in e-
organizations where hundreds of thousands of orders are placed in a commerce warehouse automation [50]. They develop a queuing net-
single day from over the globe, and multiple sellers can sell a single work model and propose a network decomposition solution method-
product, order fulfilment strategy needs to be monitored well to ensure ology for performance analysis. They mainly focus on quantifying the
effective delivery of the products to the customers promptly. Some of throughput difference between static and dynamic batching strategies
the existing research works addressing the order fulfilment problem at pick stations in a waveless order release environment. Both the
for e-commerce business has been discussed below to understand the single as well as multi-line orders have been captured in this work
current situation and recent developments and challenges in the field. for throughput performance analysis in this work. The findings exhibit
Ricker &Kalakota, in their work, identified order fulfilment as the a significant increase in system throughput with an increase in batch
hidden key that can lead to the success of an e-commerce business size. Camilleri studies various critical factors influencing the online
[47]. They have advocated that the current age customers not only service delivery of e-commerce websites, including website attractive-
buy products while shopping online but also opt for buying the service ness, website functionality, website security and consumer fulfilment
envelop alongside the product. Moreover, they focus on building a etc., in the unprecedented duration of the COVID-19 pandemic [51].
reliable relationship with support from the sellers, which is expected to Camilleri prepares a structured questionnaire for gathering data from
value the customer’s time and satisfaction. Therefore, Ricker &Kalakota online respondents from popular social media groups. The survey was
suggest that the fulfilment strategy must be the centre of focus and conducted to understand customer satisfaction with shopping websites,
the overall corporate strategy if an e-commerce organization chooses word of mouth and customer loyalty.
to compete on operational excellence and customer responsiveness. Further, Camilleri observed that the customers highly value web-
Furthermore, the competitive environment needs to be assessed while site features and order fulfilment potentiality. These factors improve
selecting an appropriate fulfilment strategy and integrating planning, customer experience and satisfaction and generate positive reviews on
warehousing and logistic processes. Effectiveness, value differentiation, social media and increase repeat sales. They strongly claim that thor-
less capital intensiveness, flexibility and channel extendibility are some ough attention to the functionality and appeal of e-commerce websites
of the factors identified by Ricker &Kalakota for selecting the fulfilment and post-sale services will increase customer satisfaction and create
strategy. They further discussed the execution of collaboration among new business opportunities.
companies that form alliances to satisfy customer demand. They also One more joint inventory and fulfilment approach has been pro-
reflect on customer-driven tailored fulfilment strategies that centre on posed by Abouelrous et al. which is again based upon a two-stage
intercompany order fulfilment models eliminating redundant processes. stochastic approximation approach [52]. They study the inventory
Moreover, since the order fulfilment process is dynamic, all potential optimization problem when retailers encounter online and offline de-
constraints, such as manufacturing, transportation, inventory, and sup- mands in peak seasons. They have adopted the Good Turing sampling
ply chain, must be considered and addressed dynamically to select a estimator framework and Linear Programming for reducing scenarios.
feasible strategy. Then they approximate the proposed two-stage stochastic optimization
Many researchers have focused on various strategies adopted for approach on these reduced scenarios. The simulation results exhibit
optimal allocation of orders for fulfilment, while others focus on mini- significant cost reduction over short time horizons to minimize the total
mizing fulfilment costs. Han et al. have considered stochastic demands cost. In another recent work on multi-warehouse collaborative distribu-
over multiple product categories while optimizing the serial supply tion route optimization, Sun & Gao have considered order splitting as
chain in their paper. They adopt a mesh adaptive direct search and a practical approach to handle the unavailability of certain goods in
focus on order splitting approach for making production and distribu- a specific warehouse [53]. They have proposed a column generation
tion decisions [48]. However, although order splitting is an appropriate approach that integrates a genetic algorithm for obtaining the optimal
fulfilment strategy, the increase in the order quantity may affect the solution. They further analyse the factors influencing cost and distri-
overall profit due to the increased fulfilment cost of the split orders, bution efficiency. Simulation results demonstrate the effectiveness of
thus making it quite irrational. the proposed approach. Next, Lei et al. studied a joint product framing

3
P. Nanda and S. Patnaik Decision Analytics Journal 7 (2023) 100227

and order fulfilment problem in the presence of both inventory and The order fulfilment process is expensive and highly complex; any
cardinality constraints in an e-commerce retailer scenario [54]. They mistake may cost the organization in terms of not only monetary loss
have considered a finite horizon for selling with no replenishment but also in terms of loss of loyal customers [14,70,71]. Therefore an
opportunity as the problem premise. They propose a two-step random- optimal strategy is most required to fulfil the orders placed by the
ized heuristic policy under the multinomial logit model based on the customers. From the previously discussed literature, some of the most
deterministic approximation approach for the stochastic control prob- commonly encountered challenges have been summarized as follows:
lem. The two-step scheme includes matrix decomposition and argument
rescaling. 1. Every day, massive amounts of orders are placed on e-commerce
Numerical experiments ensure optimal solutions even in large-scale websites. Usually, 100 thousand orders are picked up per day
practical problems. It claims to reduce shipping costs significantly. from the warehouses.
Lastly, in the recent work, Ma studies the dynamic order fulfilment 2. Maintaining a diverse range of goods has led to a constant in-
problem in a multi-item e-commerce scenario where multi-item cus- crease in the number of Stock Keeping Units (SKU), thus making
tomer orders need to be dispatched as a combination from fulfilment it difficult to maintain the balance.
centres with available inventory from the order received [55]. Ma 3. Further, this leads to increased complexity in inventory manage-
has considered finite horizon duration for inventory replenishment ment while dynamic behaviours of markets add to the uncer-
and focuses on improvisation of the deterministic Linear Programming tainty of demand forecasting strategies.
(DLP) approach proposed by Jasin and Sinha [56]. Ma improves the 4. This increased complexity and uncertainty results in frequent
correlated rounding problem identified by Jasin and Sinha [56] in the Stock-Outs.
same experimental set-up and generates a suboptimal solution by a 5. Again, shipping costs increase exponentially due to the increased
factor of 1+ln(q). Ma further presents a sparse network scheme which number of orders.
reduces the classical set cover issue and guarantees to be optimal in
terms of dependency. Ma claims that the proposed approach shortens 4. Constraints in order fulfilment process
code, improving performance and reducing runtime.
However, the most relevant works have been summarized in Some of the most commonly encountered constraints during order
Table 1. fulfilment include obtaining the most optimal weight, such as that
it should be, at most, the maximum weight allowed by the dispatch
3. Challenges in order fulfilment process vehicle. At the same time, it should be reasonable to increase the
shipping cost during the transition. The assumptions for defining the
Some of the most common challenges in an order fulfilment system constraints existing in the order fulfilment process can be summarized
range from shortages in the availability of goods to sudden surges as listed below followed by the mathematical formulation of the same.
in demand for specific goods due to festival seasons or changes in
climatic conditions. Shortage in the availability of goods may result in i. In order to avoid redundancy of picking up an SKU (Stock
stock-outs which may lead to customer dissatisfaction, thus hampering Keeping Units) from different warehouses for a given order, each
the sellers’ brand value [4,5]. Also, to avoid stock-outs, the number SKU is fulfilled only once.
of goods available on hand cannot be increased drastically since it ii. To minimize the aggregation cost of orders, transhipment be-
may also increase the cost of storage and transportation. Furthermore, tween different warehouses must be limited to once.
since the demand for goods has always been volatile, there might be iii. SKU transhipment cost of aggregated warehouses over a set of
chances that the demand for certain goods might decrease remarkably orders in a finite timeline is 0 in constraint (2).
due to uncertain reasons leading to the risk of stocking goods and iv. In any specific warehouse, the quantity of ordered products is
increasing holding costs. So, the optimal amount of inventory should less than the available inventory of the Warehouse.
be maintained by careful planning, forecasting and predicting customer
v. For any package, the total weight of sub-order packages must be
demands. Another significant challenge primarily responsible for the
less than the maximum allowed weight to be loaded.
delay in the order fulfilment process is the logistics planning limita-
vi. The total volume of the sub-order package must be less than the
tions. Late deliveries, missed deliveries, or even delivered packages
maximum volume allowed.
with damaged goods may ruin the sellers’ reputations [56,67,68].
Therefore, careful monitoring of the delivery of the goods to avoid
4.1. Mathematical modelling of the identified constraints
mishandling directly influences the satisfaction of end customers and
end-to-end performance. Other challenges arise due to external factors
The identified constraints are then modelled into mathematical
such as accidents, strikes, weather catastrophes and natural calamities
equations and integrated into the multi-agent system along with other
such as earthquakes [69]. Thus, in each stage of the order fulfilment
process, the trade-offs between the costs and benefits accrued concern- inputs through the constraint agent. At relevant breakpoints the sys-
ing various alternative operational options must be evaluated while tem will coordinate with the constraint agent to verify whether the
designing the fulfilment strategy. A poorly planned fulfilment strategy corresponding applicable constraints are satisfied or not.
may lead to the cancellation of orders and loss of money and customers. 1. To avoid redundancy
Again, sellers must provide price discounts to customers who purchase
a higher volume of goods to increase customer satisfaction. CountSKUi _fillup = 1, 𝑆𝐾𝑈𝑖 ∈ 𝑂𝑠 , 𝑖 ≤ 𝑠𝑖𝑧𝑒(𝑂𝑠 ) (1)
Additionally, during peak demand priority status of the orders to be
where
fulfilled immediately must be planned carefully so as not to dissatisfy
Os represents the number of SKUs in an order,
existing loyal customers while attempting not to lose new customers. In
SKUi represents the ith SKU in an order,
current competitive business scenarios, during peak demand periods, a
hybrid approach is adopted by the Omni-channel e-commerce retailers size(Os ) represents the count of SKU in an order
combining both in-house fulfilment approaches with outsourcing to 2.
third-party providers as well as drop shipping the bulk goods directly Count_SKU_Transshipment 𝑊𝑖, 𝑊𝑗 = 1, (2)
from manufacturers. In such cases, coalitions must be formed among
multiple sellers and third-party delivery partners to fulfil orders placed where
by the same customer to increase productivity in case of employee 𝑤𝑖 = 𝑖th warehouse, i = 1,2,. . . ,n
strikes or supply shortages at a single seller. 𝑤𝑗 = jth warehouse, j = 1,2,. . . ,n & j ≠ i

4
P. Nanda and S. Patnaik Decision Analytics Journal 7 (2023) 100227

Table 1
Summary of most relevant related works.
Observation Model Key model attributes Limitations
1. Torabi et al. [57] ∙ Introduces a ‘‘window of decision opportunity’’ approach for ∙ Demand uncertainty and
choosing the best available alternative. uncertainty in delivery lead time
∙ Optimal solutions are generated using Benders decomposition has also not been considered
approach. which is relevant to real-time
∙ Mixed-integer programming model has been adopted for scenarios.
transportation cost minimization.
2. Bhargava et al. [58] ∙ A best matching protocol (BMP) has been developed for selecting ∙ Since, the matching protocol
an order fulfilment agent. requires a centralized agent to
∙ The protocol considers pre-defined quality and cost as metrics for match demand agents to
determining the best matches. fulfilment agents, information
∙ Optimal solutions are generated through this collaboration among sharing has limitations.
the agents to reduce the fulfilment cost.
3. Leung et al. [59] ∙ A cloud based e-order fulfilment pre-processing system (CEP) has ∙ Tedious and time consuming
been proposed at the distribution centres to improve efficiency. ∙ Slow and Expensive for
∙ Genetic algorithm has been incorporated in the decision support non-linear constraints.
system for order grouping.
∙ Rule-based inference engine is adopted to generate the operating
guide lines that minimize repeated visits to storage locations.
4. Lei et al. [60] ∙ A two-heuristic control approach with an objective of maximizing ∙ Optimal solution is intractable
the total expected profits has been adopted in a joint pricing and ∙ Lacks Flexibility in uncertainty
fulfilment optimization set up.
∙ In the first stage, the heuristic utilizes a deterministic
approximation approach for the control parameters.
∙ The second stage heuristic improves the first heuristics by
adaptive adjustment to meet the order demand.
5. Luo et al. [61] ∙ A synchronized two-stage solution has been proposed for ∙ Poor Exploitation
production-warehouse management based order fulfilment process. ∙ Slow and Expensive for
∙ The first stage focuses on minimizing the processing time using non-linear constraints.
Ant Colony Optimization (ACO).
∙ The second stage focuses on minimizing warehouse space
occupation by dynamically sequencing all the products and
sub-products using genetic algorithm (GA).
6. Lu and Wang [62] ∙ An event-based model has been considered to address the ∙ Premature Convergence
multi-objective problem of supply chain partnership integration. ∙ Poor Exploitation
∙ The objectives of the multi-objective approach focuses on
maximization of product quality, supplier service and
trustworthiness while minimizing the cost incurred.
∙ A metaheuristic PSACO approach has been applied to solve the
multi-objective problem by hybridizing the Particle swarm
Optimization (PSO) and Ant Colony Optimization (ACO).
7. Zhu et al. [63] ∙ In a multi-warehouse online supermarket scenario, product ∙ For Large scale orders, splitting
category allocation is being optimized to minimize order splitting. of orders might incur higher
∙ A K-links heuristic clustering approach has been adopted for shipping cost to be bear by the
optimizing allocation of products to different warehouses with the seller.
objective of minimizing number of order splits.
∙ A binary integer programming model has been formulated for the
clustering the products.
8. Vazquez-Noguerol et al. [64] ∙ A mixed integer linear programming model has been proposed for ∙ Decision Variables are restricted
e-fulfilment strategies in warehouse-based supermarket chains to integer values.
∙ A time window-based approach has been used for picking and ∙ This limits the consideration of
transportation delivery routes. nonlinear effects
∙ It focuses on maximizing quality of service and minimizing
fulfilment cost.
9. Song et al. [65] ∙ A two-step order fulfilment strategy selection model has been ∙ Poor Exploration
proposed to reduce fulfilment cost. ∙ Lack of optimal Solution
∙ The first step uses greedy algorithm to split orders and calculate ∙ Sensitivity to errors
the fulfilment cost.
∙ The second step uses linear programming approach for merging
sub-orders.
∙ Then the split cost and merge cost is compared to identify the
lowest cost for the order fulfilment in a multi-warehouse
collaborative distribution setting.
10. Akyüz et al. [66] ∙ An iterative meta-heuristic approach combined with mixed integer ∙ Gets trapped in Local Minima
programming has been proposed for fulfilment cost minimization in
a multi-item order based Omni-channel e-commerce scenario which
generates set covering solutions.

5
P. Nanda and S. Patnaik Decision Analytics Journal 7 (2023) 100227

3. For a set of orders, O, in a finite time, T, mutually agreed upon conditions while sharing the delivery cost for a
win–win situation. Thus, the order fulfilment problem is transformed
∑
𝑖,𝑗=𝑛
Cost_TranshipmentWi,Wj = 0 (3) into a multi-objective optimization problem where optimization occurs
𝑖=𝑗=1,𝑗≠𝑖 among the three conflicting objectives, i.e., the delivery cost needs to
be minimized, and the average lead time needs to be minimized. In con-
4.
trast, the review score needs to be maximized. Therefore, in the second
(Oi Q < Ii Q) (4) phase, a set of Pareto optimal solutions is generated using the Pareto
front Method, where a set of non-dominated solutions with trade-offs
where
between distinct and conflicting objectives are generated [41–43]. The
Oi Q represents quantity of ordered products,
main goal of the nature-inspired multi-agent framework is to integrate
Ii Q represents the available inventory of the ith product in a the given historical data and evaluate the delivery cost incurred by
Warehouse. the different order fulfilment strategies to find the optimal sellers to
5. fulfil the order while minimizing the delivery cost. Integration of a
PW < Wt max (5) coalition formation approach among the competitive participants in an
e-commerce trading environment can increase collaboration support,
where leading to maximizing profit. Each order can either be fulfilled from a
PW represents the total weight of sub-order package, single seller, or it can be fulfilled from multiple sellers. Again, in the
Wtmax represents the maximum allowed weight to be loaded. case of multiple sellers, the order can either be segregated into small
6. chunks to be directly delivered to the customer from the individual
PV < Vmax, (6) warehouse or merged into packages and sent to the customers. Thus,
the problem is further divided into two sub-approaches, i.e., the order
where segregate approach and the order aggregate approach. The develop-
PV represents the total volume of sub-order package, ment of the prototype is further divided into sub-modules, the first of
Vmax represents the maximum volume allowed. which is shown in Fig. 1. The multi-agent system generates insights
The overall process of the proposed approach is schematically from past sales data, induces various observed constraints into the
shown in Fig. 1. The proposed approach consists of multiple system and then optimizes price and also monitors the performance
agents where apart from the constraint agent and order fulfilment evaluation of the system for improving the accuracy of the system.
strategy agent there are seller agents which are associated with
a particular order for which the optimal solution is being solved. 5.1. Multi-agent combined action:
The constraint agent and order fulfilment strategy agent deployed
for setting constraint criteria and cost minimization. The overall The Fig. 2 shows the underlying structure of the multi-agent system
process begins with the initialization of all the agents. Initially, responsible for the collective and cooperative action for coalition for-
each seller agent has its decision vector and priority vector where mation. It presents a constrained order fulfilment strategy for delivering
each seller agent assigns a priority weight to all other seller agents the products ordered by the customers with minimal cost involved.
whether they communicate or not. Then as the process progress The various agents involved in the combined action system include
each seller agent evaluates its objective function and negotiates (i) the source agent, (ii) the data agent, (iii) the constraint agent (iv)
with other seller agents through priority weight matching to form rank_ordering agent (v) the OFS agent, (vi) the order segregate agent
coalition. Once the seller agents agree to form coalition the non- and (vii) the order aggregate agent.
dominating solution is identified and saved in Pareto archive. Fig. 2 shows the overall workflow of the OFS Agent, which is respon-
In each iteration, the proposed approach checks for convergence sible for selecting the cost-optimal fulfilment strategy and generating
or maximum iteration reached as the stopping criteria. The de- the sequence of seller ids that must be followed for delivery of the
cision vector and priority weight vector are updated once in orders while integrating various constraints.
each iteration using gradient decision. The robustness of gradient (i) Source Agent: The source agent is responsible for loading the
descent approach is well-suited for optimizing the disagreement required CSV files into data frames for further processing.
among the seller agents in the multi-agent optimization. Once the (ii) Data Agent: The data agent is responsible for all the pre-processing
stopping criteria is met, the non-dominating solutions will be rank phases of the data in the data frames, including data cleaning, missing
ordered according to their fitness value and Pareto front solutions value treatment using forward fill and backwards fill, handling of NAN
will be returned from the Pareto archive as optimal solutions. values etc.
(iii) Constraint Agent: The constraint agent is responsible for iden-
5. The proposed multi-agent framework tifying related constraints and formulating them into policies to be
followed by other agents.
With the mathematical formulation of the identified constraints, (iv) Rank_Ordering Agent: The rank_ordering agent computes the rank
a multi-agent framework has been proposed that integrates a nature- for each seller based on the average lead time taken to fulfil orders and
inspired approach to accomplish the order fulfilment process dynami- the review scores received by them on the fulfilment of the orders.
cally. The proposed Multi-Agent Coalition-based framework optimizes (v) OFS Agent: OFS Agent is responsible for identifying the cost-optimal
the order fulfilment process following the crocodile predatory strat- fulfilment strategy and generating the sequence of seller_ids that must
egy presented by Nanda et al. [40]. In the crocodile predatory strat- be followed to deliver the orders while integrating various constraints.
egy, various traits of crocodiles have been studied and presented that (vi) Order Segregate Agent: The order segregates agent is responsible
crocodiles, even though are territorial animals, while hunting herds of for splitting an order into different parts based on the sellers with the
migrating animals they, form temporary alliances or coalitions with availability of the ordered item and calculates the transportation cost
other competitors for mutual benefits. The cooperative predatory strat- of fulfilling the order separately.
egy of crocodiles has been modelled using the probability mass function (vii) Order Aggregate Agent: The aggregate order agent combines
following the Skellam distribution, which gives the discrete probability products from different sellers into single packages and calculates the
distribution of the difference between two statistically independent transportation cost for the aggregated fulfilment strategy.
random variables. In the current research work, the proposed frame- As shown in Fig. 3, the order segregate agent follows a split ap-
work forms a coalition among sellers, although being competitors with proach that divides an order into different parts based on the seller

6
P. Nanda and S. Patnaik Decision Analytics Journal 7 (2023) 100227

Fig. 1. Flowchart of the overall process of the proposed approach.

with the availability of the ordered item. Once the order is divided each strategy. The delivery cost includes both the transportation cost
into parts, it can be either directly sent from the seller to the customer as well as the transhipment cost of the ordered items.
or be transhipped to a fulfilment centre and combined into a single The aggregate order agent also partitions the order into several
package and addressed to the individual customer who has placed the parts while checking stock availability at different seller warehouses.
order. Then, each fulfilment centre sends the package to the addressed It computes the delivery cost of each fulfilment centre, including the
customer. However, the choice of the fulfilment strategy, whether to transhipment cost between seller warehouses and the transportation
adopt an order segregation strategy or an order aggregation strategy, cost from the fulfilment centre to the customer. A weight has been
is done by the OFS agent based on the minimization of delivery cost by assigned to the delivery cost to evaluate the alternatives available and

7
P. Nanda and S. Patnaik Decision Analytics Journal 7 (2023) 100227

Fig. 2. Constrained order fulfilment strategy.

Fig. 3. OFS agent workflow.

8
P. Nanda and S. Patnaik Decision Analytics Journal 7 (2023) 100227

rank them in order of dominance. The weight has been assigned to the At any iteration k, the priority weight vector and decision vector of
sellers based on lead time. The lead time is the duration taken from agent i using
when an order is placed to when the good is delivered to the end
∑
𝑛
customer and is given by the number of days. The lowest lead time 𝑥𝑖 (𝑘 + 1) = 𝑎𝑖𝑗 (𝑘)𝑥𝑗 (𝑘) − 𝛼𝑖 (𝑘)∇𝑓𝑖 𝑥𝑖 (𝑘) (12)
taken by the seller dominates in the rank ordering over the sellers with 𝑗=1
higher lead times. Thus, the alternative having minimal delivery cost where xi (0) = (xi1 (0), . . . , xin (0)) represents the initial decision vector
along with minimal lead time will be considered the optimal option and of agenti , aij is the priority weight assigned by the agenti to agent j , 𝛼i
presented to the OFS agent. is the step size of agenti and ∇fi xi (k) is the gradient vector of agent i
The OFS agent then compares the delivery costs provided by the at xi (k).
order segregate agent and the aggregate order agent to select the
fulfilment strategy with a minimum delivery cost.
5.2. Mathematical notation for OFS agent
Assumptions
Mathematically, Sk can be further represented as
(i) Based on the above discussion, the following assumptions have
been considered. ⎧𝑆 𝑖𝑓 𝐶𝑇 𝑟𝑎𝑛𝑠𝑆𝑒𝑔 < 𝐶𝑇 𝑟𝑎𝑛𝑠𝐴𝑔𝑔
(ii) Multiple fulfilment centres can be present in the same city with the ⎪ 𝑠𝑒𝑔,
𝑆𝑘 = ⎨ (13)
capability of packaging and merging orders from SKUs. ⎪𝑆𝐴𝑔𝑔, 𝑖𝑓 𝐶𝑇 𝑟𝑎𝑛𝑠𝑎𝑔𝑔 < 𝐶𝑇 𝑟𝑎𝑛𝑠𝑠𝑒𝑔
(iii) Once the order is placed, the fulfilment centre’s location and the ⎩
customer shipping location are known. where 𝐶𝑇 𝑟𝑎𝑛𝑠𝑆 = Transportation cost incurred while segregating the
𝑆𝑒𝑔
(iv) The volume and weight of each ordered item must be, at most, the
participant seller warehouses to deliver the products directly to the
maximum allowed capacity for the package.
customer.
(v) Transportation cost is involved when the package is shipped from
And 𝐶𝑇 𝑟𝑎𝑛𝑠𝑆 = Transportation cost incurred while aggregating the
the fulfilment centre to the customer, while transhipment cost is 𝐴𝑔𝑔
products of the order from the participant seller warehouses to deliver
involved when a package is shipped from one fulfilment centre to
them as a bundle to the customer.
another.
(vi) Cost (charge) of a vehicle is fixed for all instances.
5.3. Mathematical notation for order aggregate agent
The goal of the OFS agent is to find the optimal fulfilment strategy,
not only minimizing the transportation cost but also identifying the The objective of the Order Aggregate Agent:
sellers with the minimal average lead time taken for completing the 𝑛
order and maximum review scores. Since, here, the optimal solution 𝑂𝑏𝑗𝑒𝑐𝑡𝑖𝑣𝑒 ∶ 𝑌 = 𝐶𝑇 𝑟𝑎𝑛𝑠𝑆 = min(𝐶𝑇 𝑟𝑎𝑛𝑠𝐶𝑊 ) (14)
𝐴𝑔𝑔 𝑖=1 𝑖
of the first and second objective functions, i.e., the transportation cost
s.t. constraints (1)–(6)
and the average lead time, needs to be minimized, which might be
conflicting and not coincide with the objective of the third objective where 𝐶𝑇 𝑟𝑎𝑛𝑠𝐶𝑊 = Transportation cost for ith combination of partici-
𝑖
function, i.e., the review score needs to be maximized. Therefore, here, pant seller warehouses
we have considered Pareto-optimal solutions, i.e., those that are non-
∑
𝑛
dominated by any other feasible solution, have been considered the Which can be calculated as, 𝐶𝑇 𝑟𝑎𝑛𝑠𝐶𝑊 = D
𝑖
solution set. Hence, the overall objective function can be formulated 𝑖=1 𝑊𝑖 𝐷𝐶𝑛𝑒𝑎𝑟𝑒𝑠𝑡
as a multi-objective optimization problem as follows: × (𝐶𝑝𝑖 + 𝐶𝑣𝑖 ) (15)
Min(Y) = f(x) = (f1 (x), f2 (x)) (7) where 𝐷𝑊𝑖 𝐷𝐶𝑛𝑒𝑎𝑟𝑒𝑠𝑡 = Distance between the seller warehouse i and the
and Max(Y) = f(x) = (f3 (x)) (8) identified nearest distribution centre,
Subject to: gi (x) ≥ 0, i = 1, 2, … , 𝑘 𝐶𝑝 = unit path cost,
hi (x) ≥ 0, i = 1, 2, … , 𝑘 𝐶𝑣 = Vehicle cost
Further, it can be elaborated as below: 𝐷𝐶𝑛𝑒𝑎𝑟𝑒𝑠𝑡 = min(𝐷(𝐿𝑜𝑐𝐷𝐶 𝑖 , 𝐿𝑜𝑐𝐶𝑢𝑠𝑡𝑜𝑚𝑒𝑟 ))
The first objective function is to minimize the transportation cost Where 𝐿𝑜𝑐𝐷𝐶 𝑖 = Location of Distribution centrei
for order i, given by, 𝐶𝑇 𝑟𝑎𝑛𝑠𝑖 𝐿𝑜𝑐𝐶𝑢𝑠𝑡𝑜𝑚𝑒𝑟 = Location of Customer
𝑛
𝑓 1 = min (𝐶𝑇 𝑟𝑎𝑛𝑠𝑖 ) (9)
𝑖=1 5.4. Mathematical notation for order segregate agent
Subject to: CTransi (x) ≥ 0, i = 1, 2, …, n
The objective of the Order Segregate Agent can be mathematically
The second objective function is to minimize the average lead time
represented as :
for each seller j, given by, 𝐿𝑇𝐴𝑉 𝐺𝑗
∑
𝑛
𝑠 𝑂𝑏𝑗𝑒𝑐𝑡𝑖𝑣𝑒 ∶ 𝑌 = 𝐶𝑇 𝑟𝑎𝑛𝑠𝑆 = 𝐷(𝐿𝑜𝑐𝑊𝑖 , 𝐿𝑜𝑐𝐶𝑢𝑠𝑡𝑜𝑚𝑒𝑟 ) × (𝐶𝑝𝑖 + 𝐶𝑣𝑖 ) (16)
𝑓2 = min (𝐿𝑇𝐴𝑉 𝐺𝑗 ) (10) 𝑆𝑒𝑔
𝑖=1
𝑗=1

Subject to: LTAVGj (x) ≥ 0, j = 1, 2, …, s where 𝐷(𝐿𝑜𝑐𝑊𝑖 , 𝐿𝑜𝑐𝐶𝑢𝑠𝑡𝑜𝑚𝑒𝑟 ) = Distance between the location of the ith
seller warehouse and customer.
The third objective function is to maximize the review score for each
seller j, given by, 𝑅𝑆𝑗 𝐶𝑝 = unit path cost,
𝑠 𝐶𝑣 = Vehicle cost,
𝑓3 = max(𝑅𝑆𝑗 ) (11)
𝑗=1 𝐿𝑜𝑐𝑊𝑖 = Location of ith seller warehouse
Subject to: 𝑅𝑆 𝑗 (x) ≥ 0, j = 1, 2, …, s 𝐿𝑜𝑐𝐶𝑢𝑠𝑡𝑜𝑚𝑒𝑟 𝑗 = Location of Customer

9
P. Nanda and S. Patnaik Decision Analytics Journal 7 (2023) 100227

6. Dataset considered sellers fulfilling those orders were extracted from different datasets and
merged into a single data frame concerning the corresponding order ids.
• The dataset considered for the development of the prototype is Similarly, for various calculations, relevant features are extracted and
the real-time transaction data of a Brazilian e-commerce firm. merged into data frames for ease of computation.
• The dataset presents information about around 100k orders over The constraint functional agent sets the constraint criteria as dis-
two years, from 2016 to 2018, made from different places in cussed in equations and continuously monitors it while the segregation
Brazil. and aggregation costs are computed. While the Rank_Ordering func-
• This dataset presents an extensive collection of features that tional agent then computes the rank of each seller based on the review
allows viewing orders from different dimensions such as order sta- scores obtained by the sellers in fulfilling an order and the average lead
tus, freight performance, product price, mode of payment, freight time taken by the sellers to fulfil the order. Here, the review scores
cost, seller location to customer location and various product provided by the real-time data exhibit different levels of customer sat-
attributes. isfaction that range from 1 to 5. Again for this computational purpose,
• The dataset also provides a geo-location dataset with Brazilian zip we have considered those sellers whose orders have received a score
codes and coordinates of latitude and longitude. equal to or above 3. Also, an average lead time of ten days has been
• Since the real-time dataset has been used, the actual Ids for sell- considered an idle value by the rank_ordering agent for each seller for
ers, customers, products and orders have been used to maintain computational purposes. Although the average quoted lead time for
a proper synchronization. the orders is around twenty-five to thirty days, most orders have been
fulfilled with an average lead time of 10–15 days. Therefore, ten days
7. Results and discussions have been considered as an optimal lead time for rank computation.
The OFS functional agent coordinates the synchronization of up-
This research work contributes to the existing literature by provid- dated information and states among the constraint agent, rank_ordering
ing an optimal solution to the order fulfilment process with various agent, cost segregate agent and cost aggregate agent. However, the
control policies to enhance the performance criteria. The proposed constraint agent and rank_ordering agents are monitored continuously
multi-agent-based coalition formation framework for the order ful- by the OFS functional agent in regular intervals for each time the cost
filment process has been tested on real-time data to optimize order segregate and cost aggregate agents are invoked. Firstly, the orders are
fulfilment cost across multiple dimensions. The main objective of the checked for whether they need a single product or multiple products to
proposed framework is to improve the agility and productivity of the fulfil it. Then, for those cases where an order needs multiple products
fulfilment process of multiple orders by identifying optimal sellers to be fulfilled, it is further checked whether a single seller or multiple
across different distribution locations by order ranking them on the sellers can fulfil it. In case there are multiple products which different
prior discussed criteria with appropriate adjustments. This framework sellers can fulfil, the cost segregate agent computes the cost of each
is cost-effective while ensuring productivity and quality of the order fulfilment by splitting the order into several sellers involved and the
fulfilment process. The framework includes initiation of the agents, co- cost required from that seller to the customer placing the order.
ordination of the constraint policies, coordination of the cost evaluating Meanwhile, the aggregate cost agent computes the fulfilment cost
agents, information sharing policies and dynamic coalition formation by aggregating all the products at the seller with a minimum freight
strategy based on specific criteria. The agents’ performance is evaluated charge. Apart from a distance, the rank of the seller generated by the
by considering a different combination of constraint policies, lead rank-ordering agent is also considered while calculating aggregation
times and review scores provided by the customers. The tardiness in cost. Next, the segregate and aggregate costs are compared, and the
the fulfilment process was observed by incorporating uncertainty by OFS functional agent considers the minimum. With the change in the
observing the orders’ lead times and expected due dates. fulfilment cost of the orders, the fulfilment strategy chosen by the
In order to validate the proposed multi-agent framework (MAS- OFS functional agent also varies. For those cases where the customer
CPS), several instances with different configurations of fulfilment costs, location is too close to that of the seller fulfilling the orders, the
lead time and customer reviews have been considered with real-time cost segregation strategy is a better choice than the cost aggregation
data. The Brazilian dataset spanning a horizon of around two years strategy. However, a cost aggregation strategy is preferred for orders
consists of 3095 different sellers distributed all over Brazil, selling with a large number of products. Sometimes, coalitions must be formed
about 32 951 products belonging to 73 categories broadly and respon- among the sellers, even after being competitors, to generate a cost-
sible for fulfilling around 99 441 orders for 96 097 customers spread effective strategy. The OFS agent then generates the ids of the optimal
across Brazil. The MAS-CPS framework is initiated by loading the sellers that can fulfil the order efficiently and cost-effectively. The
potential datasets required for analysis. The dataset has been loaded segregation cost and aggregation cost of a subset of orders have been
by the source functional agent of the framework, which is responsible represented in Fig. 4 to provide a comparative view.
for reading the ‘CSV’ files. The functional data agent performed the Again, the rank ordering of the sellers, along with the optimal values
pre-processing stages, such as cleaning the data and missing value of the segregation cost, aggregation cost, lead time and review score us-
treatment using forward and backward fills. Once the pre-processing ing the MAS-CPS, can be seen in the following table. Table 2, generated
steps are over, the functional data agent provides information about from the implementation of the proposed approach, showcases the top
the relationship between relevant features required for the computation six sellers and their respective attribute values. It can be observed that
and analysis. As discussed previously, a strong negative correlation the seller ranked as number one has an optimal average lead time of
can be observed between the lead time and review score provided by 9 days and an optimal review score of 3. While the average lead time
the customers. Next, once the relationships between the independent needs to be minimized and the review score needs to be maximized,
and dependent variables have been identified, required features from the maximum value for the review score is five. The trade-off between
various datasets are filtered and merged into different data frames for the review score and average lead time needs to be balanced while
easy computation. For example, order_ids, order purchase date, order minimizing the fulfilment cost for rank-ordering the sellers.
delivery date and estimated order delivery date, and the customer is Similarly, for the seller at rank order two, it can be seen that
placed are extracted from the order dataset. Then these features are the average lead time decreases to eight days while the review score
merged into a single data frame with other features such as product increases to value 4, which again exhibits the negative correlation
ids, quantity, seller providing these products, price of the products between the average lead time and review score. Next, variation can
and freight charge values assigned to each order concerning order ids be observed in the case of the seller rank-ordered to position four.
from another dataset. The review score for different orders and the Here, the average lead time varies with different customers due to the

10
P. Nanda and S. Patnaik Decision Analytics Journal 7 (2023) 100227

Table 2
Rank-ordering of optimal sellers.
Observation Order ID Seller ID Freight value Order segregation cost Order aggregation cost Average lead time Review score No. of items Rank (Sorted)
1. 500010 40009 6.1 21.34 3.05 9 3 3 1
2. 500055 40009 12.19 21.34 3.05 9 3 3 1
3. 500080 40040 20.08 33.13 13.05 8 4 2 2
4. 500124 40057 20.16 36.52 16.36 9 4 2 3
5. 500158 40110 36.19 72.38 36.19 9 5 2 4
6. 500165 40110 13.77 27.54 13.77 7 4 2 4
7. 500188 40110 13.77 27.54 13.77 7 4 2 4
8. 500235 40110 44.32 88.65 44.32 7 5 2 4
9. 500270 40110 137.12 274.25 137.12 9 5 2 4
10. 500310 40110 8.45 36.6 8.45 8 4 2 4
11. 500345 40151 20.99 41.98 20.99 7 3 2 5
12. 500379 40211 29.2 58.4 29.2 9 3 2 6
13. 500410 40211 29.2 58.4 29.2 9 3 2 6

Fig. 4. Segregation cost vs. aggregation cost.

location factor, which is also responsible for the fluctuation in average

lead time. Therefore, a set of optimal values has been generated, and
the dominating solution has been considered. Similarly, the rest of
the sellers are subjected to the proposed MAS-CPS approach for rank-
ordering while optimizing the average lead time, review score, and
fulfilment cost.
The Pareto Optimal solution generated in Table 2 is the key finding
of the proposed MAS-CPS approach, where each agent has a specific
function assigned to it. Due to conflicting objectives, each agent com-
petes and collaborates with other agents to explore the Pareto Optimal
solutions using the Pareto Front Method. Since most of the existing
literature deals with the order fulfilment process using mixed integer
linear programming or deterministic approximation approaches, it is
not easy to find comparative results. However, the fitness function’s
convergence for cost optimization is shown in Fig. 5. Fig. 5. Comparative performance analysis of MAS-CPS.
The fitness value of the objective function for cost minimization has
been presented along the 𝑦-axis, while the number of iterations has
been presented along the 𝑥-axis. A total of 100 iterations have been 8. Conclusion
considered in this experimental setup, along with other parameters. It
presents the comparative convergence diagram for the proposed MAS- This paper proposed a distributed multi-agent framework to op-
CPS approach, which has been compared with a hybrid metaheuristic timize the order fulfilment process and identify the optimal set of
approach PSO_ACO combining Particle Swarm Optimization and Ant sellers that fulfil an order reliably. The proposed approach allows the
Colony Optimization [62]. Thus, it can be seen in Fig. 4. that PSO_ACO exploration of the Pareto Front of the overall objectives influencing the
suffers premature convergence by converging to stability too soon to final results of the agents. After identifying the constraints involved
minimize cost at around five iterations. At the same time, the pro- in the order fulfilment problem, the main objective of optimizing the
posed MAS-CPS approach exhibits better exploitation. Therefore, the process is to minimize the fulfilment cost. Two more factors associated
proposed MAS-CPS strategy attempts to re-model the Order Fulfilment with successfully fulfilling e-commerce orders have been considered:
Strategy to attain a certain level of autonomy by evaluating and op- the average lead time and the review score. The sellers are subjected
timizing the interdependent factors by adopting a multi-agent system to a rank-ordering approach proposed by the MAS-CPS strategy consid-
and defining separate functional agents to accomplish specific tasks. ering minimal cost, average lead time, and maximum possible review

11
P. Nanda and S. Patnaik Decision Analytics Journal 7 (2023) 100227

score. Hence, the order fulfilment problem has been formulated into [20] A. Al-Khayyal, M. Alshurideh, B. Al Kurdi, A. Aburayya, The impact of electronic
a multi-objective problem considering the above-discussed objectives. service quality dimensions on customers’e-shopping and e-loyalty via the impact
of e-satisfaction and e-trust: A qualitative approach, Int. J. Innov. Creat. Change
The effectiveness of the proposed approach has been demonstrated
14 (9) (2020) 257–281.
on real-time benchmark data. It exhibits that the proposed approach [21] J. Acimovic, S.C. Graves, Making better fulfillment decisions on the fly in an
presents a set of Pareto-optimal solutions to the framework for select- online retail environment, Manuf. Serv. Oper. Manag. 17 (1) (2015) 34–51.
ing the order fulfilment strategy that best suits the requirement. The [22] L. DeValve, Y. Wei, D. Wu, R. Yuan, Understanding the value of fulfillment
dynamic pricing strategy and inventory replenishment processes can flexibility in an online retailing environment, Manuf. Serv. Oper. Manag. (2021).
[23] L. Sun, G. Lyu, Y. Yu, C.P. Teo, Fulfillment by Amazon versus fulfillment by
be integrated into the proposed MAS-CPS framework as a future re- seller: An interpretable risk-adjusted fulfillment model, Nav. Res. Logist. 67 (8)
search direction. Also, this framework is continuously being improved (2020) 627–645.
to present it as a full fledge distributed multi-objective optimization [24] A. Johnson, S. Carnovale, J.M. Song, Y. Zhao, Drivers of fulfillment performance
technique by testing it over benchmark functions. in mission critical logistics systems: An empirical analysis, Int. J. Prod. Econ. 237
(2021) 108138.
[25] D.A. Goldberg, D.A. Katz-Rogozhnikov, Y. Lu, M. Sharma, M.S. Squillante,
Declaration of competing interest Asymptotic optimality of constant-order policies for lost sales inventory models
with large lead times, Math. Oper. Res. 41 (3) (2016) 898–913.
The authors declare that they have no known competing finan- [26] K.J. Roodbergen, I.B. Kolman, Order fulfillment and logistics considerations
for multichannel retailers, in: Logistics and Supply Chain Innovation, Springer,
cial interests or personal relationships that could have appeared to
Cham, 2016, pp. 183–196.
influence the work reported in this paper. [27] Y. Zhu, A. Deshpande, B. Quanz, L. Cao, A. Koc, X. Liu . . . , R. Tan, Improving e-
commerce order fulfillment for peak times via incorporating fulfillment network
Data availability load balancing, in: IIE Annual Conference. Proceedings, Institute of Industrial
and Systems Engineers (IISE), 2017, pp. 1806–1810.
[28] K.H. Leung, S.W. Cheng, K.L. Choy, D.W. Wong, H.Y. Lam, Y.Y. Hui . . . ,
The authors do not have permission to share data V. Tang, A process-oriented warehouse postponement strategy for e-commerce
order fulfillment in warehouses and distribution centers in Asia, in: Manage-
rial Strategies and Solutions for Business Success in Asia, IGI Global, 2017,
References
pp. 21–34.
[29] J. Acimovic, V.F. Farias, The fulfillment-optimization problem, in: Operations
[1] M.J. North, C.M. Macal, Managing Business Complexity: Discovering Strategic Research & Management Science in the Age of Analytics, INFORM, 2019, pp.
Solutions with Agent-Based Modeling and Simulation, Oxford University Press, 218–237.
2007. [30] N.A. Agatz, M. Fleischmann, J.A. Van Nunen, E-fulfillment and multi-channel
[2] A. Martel, W. Klibi, Designing Value-Creating Supply Chain Networks, Springer, distribution–A review, European J. Oper. Res. 187 (2) (2008) 339–356.
Cham, 2016. [31] C.W. Chan, V.F. Farias, Stochastic depletion problems: Effective myopic policies
[3] A.H. Mohamad, G.F. Hassan, A.S. Abd Elrahman, Impacts of e-commerce on for a class of dynamic optimization problems, Math. Oper. Res. 34 (2) (2009)
planning and designing commercial activities centers: A developed approach, 333–350.
Ain Shams Eng. J. 13 (4) (2022) 101634. [32] T. Aslam, H.N. Amos, Multi-objective optimization for supply chain management:
[4] A. Bask, M. Lipponen, M. Tinnilä, E-commerce logistics: a literature research A literature review and new development, in: 2010 8th International Conference
review and topics for future research, Int. J. E-Serv. Mob. Appl. (IJESMA) 4 (3) on Supply Chain Management and Information, IEEE, 2010, pp. 1–8.
(2012) 1–22. [33] W. Alshabi, S. Ramaswamy, M. Itmi, H. Abdulrab, Coordination, cooperation
[5] E. Colla, P. Lapoule, E-commerce: exploring the critical success factors, Int. J. and conflict resolution in multi-agent systems, in: Innovations and Advanced
Retail Distrib. Manag. 40 (11) (2012) 842–864. Techniques in Computer and Information Sciences and Engineering, Springer,
[6] N. Turner, J. Aitken, C. Bozarth, A framework for understanding managerial Dordrecht, 2007, pp. 495–500.
responses to supply chain complexity, Int. J. Oper. Prod. Manag. (2018). [34] S. Berger, C. Bierwirth, Solutions to the request reassignment problem in
[7] R.E. Bawack, S.F. Wamba, K.D.A. Carillo, S. Akter, Artificial intelligence in E- collaborative carrier networks, Transp. Res. E 46 (5) (2010) 627–638.
Commerce: a bibliometric study and literature review, Electron. Mark. 32 (1) [35] E.H. Durfee, Coordination of Distributed Problem Solvers, Vol. 55, Springer
(2022) 297–338. Science & Business Media, 2012.
[8] N.V. Florea, C.A. Ionescu, M.C. Duică, S. Căpuşneanu, L. Paschia, S.G. Stanescu, [36] A. Elomri, A. Ghaffari, Z. Jemai, Y. Dallery, Coalition formation and cost
M.D. Coman, Trends and perspectives of Romanian E-commerce sector based on allocation for joint replenishment systems, Prod. Oper. Manage. 21 (6) (2012)
mathematical simulation, Electronics 11 (15) (2022) 2295. 1015–1027.
[9] J. Żurek, E-commerce influence on changes in logistics processes, LogForum 11 [37] S. Creemers, G. Woumans, R. Boute, J. Beliën, Tri-vizor uses an efficient
(2) (2015) 129–138. algorithm to identify collaborative shipping opportunities, Interfaces 47 (3)
[10] P. Fernández Campos, P. Trucco, L. HuacchoHuatuco, Managing structural and (2017) 244–259.
dynamic complexity in supply chains: insights from four case studies, Prod. Plan. [38] S.B. Jouida, S. Krichen, W. Klibi, Coalition-formation problem for sourcing
Control 30 (8) (2019) 611–623. contract design in supply networks, European J. Oper. Res. 257 (2) (2017)
[11] P. Dabidian, U. Clausen, E. Denecke, An investigation of behavioural and 539–558.
structural characteristics of CEP service providers and freight demand considering [39] Y. Rizk, M. Awad, E.W. Tunstel, Decision making in multi-agent systems: A
e-commerce in Germany, Transp. Res. Procedia 14 (2016) 2795–2804. survey, IEEE Trans. Cogn. Dev. Syst. 10 (3) (2018) 514–529.
[12] J.D. Ser, M.N. Bilbao, C. Perfecto, S. Salcedo-Sanz, A harmony search approach [40] P. Nanda, S. Patnaik, S. Patnaik, Modelling of multi-agent coordination using
for the selective pick-up and delivery problem with delayed drop-off, in: crocodile predatory strategy, Int. J. Reason.-Based Intell. Syst. 8 (3–4) (2016)
Harmony Search Algorithm, Springer, Berlin, Heidelberg, 2016, pp. 121–131. 119–131.
[13] A. Kawa, Supply chains of cross-border e-commerce, in: Asian Conference [41] P. Lindroth, M. Patriksson, A.B. Strömberg, Approximating the Pareto optimal
on Intelligent Information and Database Systems, Springer, Cham, 2017, pp. set using a reduced set of objective functions, European J. Oper. Res. 207 (3)
173–183. (2010) 1519–1534.
[14] J. Acimovic, S.C. Graves, Mitigating spillover in online retailing via [42] M. Ancău, C. Caizar, The computation of Pareto-optimal set in multicriterial
replenishment, Manuf. Serv. Oper. Manag. 19 (3) (2017) 419–436. optimization of rapid prototyping processes, Comput. Ind. Eng. 58 (4) (2010)
[15] A. Kumar, R. Chopra, R.R. Saxena, An efficient algorithm to solve transshipment 696–708.
problem in uncertain environment, Int. J. Fuzzy Syst. 22 (8) (2020) 2613–2624. [43] L.M. Torres-Treviño, F.A. Reyes-Valdes, V. López, R. Praga-Alejo, Multi-objective
[16] Y. Du, X. Hu, K. Vakil, Systematic literature review on the supply chain agility optimization of a welding process by the estimation of the Pareto optimal set,
for manufacturer and consumer, Int. J. Consumer Stud. 45 (4) (2021) 581–616. Expert Syst. Appl. 38 (7) (2011) 8045–8053.
[17] H. Wang, R. Du, T. Olsen, Feedback mechanisms and consumer satisfaction, [44] M. Gobbi, F. Levi, G. Mastinu, G. Previati, On the analytical derivation of the
trust and repurchase intention in online retail, Inf. Syst. Manag. 35 (3) (2018) Pareto-optimal set with applications to structural design, Struct. Multidiscip.
201–219. Optim. 51 (3) (2015) 645–657.
[18] S. Sharifi, Examining the impacts of positive and negative online consumer re- [45] H. Jafaryeganeh, M. Ventura, C.G. Soares, Application of multi-criteria decision
views on behavioral intentions: Role of need for cognitive closure and satisfaction making methods for selection of ship internal layout design from a Pareto optimal
guarantees, J. Hosp. Mark. Manag. 28 (4) (2019) 397–426. set, Ocean Eng. 202 (2020) 107151.
[19] X. Lin, X. Wang, N. Hajli, Building e-commerce satisfaction and boosting sales: [46] P.J. Reaidy, A. Gunasekaran, A. Spalanzani, Bottom-up approach based on Inter-
The role of social commerce trust and its antecedents, Int. J. Electron. Commer. net of Things for order fulfillment in a collaborative warehousing environment,
23 (3) (2019) 328–363. Int. J. Prod. Econ. 159 (2015) 29–40.

12
P. Nanda and S. Patnaik Decision Analytics Journal 7 (2023) 100227

[47] F. Ricker, R. Kalakota, Order fulfillment: the hidden key to e-commerce success, [60] Y. Lei, S. Jasin, A. Sinha, Joint dynamic pricing and order fulfillment for
Supply Chain Manag. Rev. 11 (3) (1999) 60–70. e-commerce retailers, Manuf. Serv. Oper. Manag. 20 (2) (2018) 269–284.
[48] X. Han, Y. Wang, S. Zhao, Optimization for serial supply chain’s control strategy [61] H. Luo, X. Yang, X.T. Kong, A synchronized production-warehouse management
based on mesh adaptive direct search and simulation, Comput. Integr. Manuf. solution for reengineering the online-offline integrated order fulfillment, Transp.
Syst. 22 (05) (2016) 1339–1346. Res. E 122 (2019) 211–230.
[49] Y. Zhao, X. Wang, L. Xin, Multi-Item Online Order Fulfillment in a Two-Layer [62] Z. Lu, H. Wang, An event-based supply chain partnership integration using a
Network, Chicago Booth Research Paper, (20-41), 2020. hybrid particle swarm optimization and ant colony optimization approach, Appl.
[50] V. Bansal, D. Roy, J.A. Pazour, Performance analysis of batching decisions Sci. 10 (1) (2019) 190.
in waveless order release environments for e-commerce stock-to-picker order [63] S. Zhu, X. Hu, K. Huang, Y. Yuan, Optimization of product category allocation
fulfillment, Int. Trans. Oper. Res. 28 (4) (2021) 1787–1820. in multiple warehouses to minimize splitting of online supermarket customer
[51] M.A. Camilleri, E-commerce websites, consumer order fulfillment and after-sales orders, European J. Oper. Res. 290 (2) (2021) 556–571.
service satisfaction: The customer is always right, even after the shopping cart [64] M. Vazquez-Noguerol, J.A. Comesaña-Benavides, S. Riveiro-Sanroman, J.C.
check-out, J. Strategy Manag. (2021). Prado-Prado, A mixed integer linear programming model to support e-fulfillment
[52] A. Abouelrous, A.F. Gabor, Y. Zhang, Joint inventory and fulfilment optimization strategies in warehouse-based supermarket chains, CEJOR Cent. Eur. J. Oper.
for an omnichannel retailer: A stochastic optimization approach, Comput. Ind. Res. (2021) 1–34.
Eng. (2022) http://dx.doi.org/10.2139/ssrn.4110440. [65] Shuanjun Song, Longguang Peng, Yuanyi Meng, Sheng Hu, Order fulfillment
[53] Y. Sun, G. Gao, Optimize the multi-warehouse collaborative distribution route decision under multiwarehouse collaborative delivery, Math. Probl. Eng. 2021
that considers the splitting of E-commerce orders, Front. Econ. Manag. 3 (5) (2021) 6663416, 12 pages.
(2022) 357–369. [66] M.H. Akyüz, İ. Muter, G. Erdoğan, G. Laporte, Minimum cost delivery of
[54] Y. Lei, S. Jasin, J. Uichanco, A. Vakhutinsky, Joint product framing multi-item orders in e-commerce logistics, Comput. Oper. Res. 138 (2022)
(display, ranking, pricing) and order fulfillment under the multinomial logit 105613.
model for e-commerce retailers, Manuf. Serv. Oper. Manag. 24 (3) (2022) [67] N. Buchbinder, T. Kimbrelt, R. Levi, K. Makarychev, M. Sviridenko, Online
1529–1546. make-to-order joint replenishment model: primal dual competitive algorithms,
[55] W. Ma, Simple and order-optimal correlated rounding schemes for multi-item in: Proceedings of the Nineteenth Annual ACM-SIAM Symposium on Discrete
E-commerce order fulfillment, 2022, arXiv preprint arXiv:2207.04774. Algorithms, 2008, pp. 952–961.
[56] S. Jasin, A. Sinha, An LP-based correlated rounding scheme for multi-item [68] C. Isac, E-fulfillment-a new challenge for electronic business, Ann. Univ.
e-commerce order fulfillment, Oper. Res. 63 (6) (2015) 1336–1351. Petroşani Econ. 14 (2014) 121–128.
[57] S.A. Torabi, E. Hassini, M. Jeihoonian, Fulfillment source allocation, inventory [69] F. Fernández-Bonilla, C. Gijón, B. De la Vega, E-commerce in Spain: Determining
transshipment, and customer order transfer in e-tailing, Transp. Res. E 79 (2015) factors and the importance of the e-trust, Telecommun. Policy 46 (1) (2022)
128–144. 102280.
[58] R. Bhargava, R.R. Levalle, S.Y. Nof, A best-matching protocol for order fulfillment [70] A. Vargas, S. Patel, D. Patel, Towards a business model framework to increase
in re-configurable supply networks, Comput. Ind. 82 (2016) 160–169. collaboration in the freight industry, Logistics 2 (4) (2018) 22.
[59] K.H. Leung, K.L. Choy, P.K. Siu, G.T. Ho, H.Y. Lam, C.K. Lee, A B2C e-commerce [71] J.M. Andrews, V.F. Farias, A.I. Khojandi, C.M. Yan, Primal–dual algorithms for
intelligent system for re-engineering the e-order fulfilment process, Expert Syst. order fulfillment at urban outfitters, inc, INFORMS J. Appl. Anal. 49 (5) (2019)
Appl. 91 (2018) 386–401. 355–370.

13