As a result, the parcels take over many of the decision-making abilities of humans

THE IMPACT OF DECENTRALISED INTELLIGENCE IN THE PARCEL DELIVERY INDUSTRY: A FIELD EXPERIMENT

Master thesis Technology and Operations Management & Supply Chain Management Faculty of Economics and Business, University of Groningen

24 June 2019

Rosemarie Cramer Student number: 2373548

Supervisor:

dr. ir. P. Buijs Second assessor:

dr. O.A. Kilic

2 ABSTRACT

A shift towards a more decentralised planning system is often proposed as a solution to the growing complexity of the parcel delivery industry. The parcels in the proposed system are intelligent and are able to collect and process information based on which they make decisions regarding their routes. As a result, the parcels take over many of the decision-making abilities of humans. This thesis is the first to empirically study the effects of decentralised intelligence. We test its impact on the efficiency of working days of parcel delivery vehicle drivers.

A field experiment is conducted among two parcel delivery vehicle drivers who assign their scanning and loading tasks to depot operators. This opens up possibilities to automatically load vehicles, which means the drivers are available for longer tours. On the other hand, this possibly reduces the tacit knowledge of drivers on where the parcels are located in their vehicles, reducing the efficiency of the tour as a result of higher search time. So, in return, the drivers are provided with external information which resembles the decentralised intelligence coming from the parcels. We measure whether the search time, the time drivers need to find the parcel at the addresses of delivery, would increase under different levels of external information.

The findings suggest that, compared to self-loading, drivers barely use extra search time under the loading policy by which they have more accurate external information. Therefore, by (1) letting drivers spend more time on the road by assigning the loading task to depot operators, (2) providing sufficient external information, and (3) taking into account the volume capacity of the vehicles, this thesis suggests a reduction of approximately 20% of vehicles on the road.

Overall, this study contributes to the understanding of the impact of parcel vehicle drivers’

situational awareness loss and to directions in overcoming that impact, the types of information needed from a decentralised intelligent system, the knowledge concerning human involvement in decentralised intelligence, and the role of decentralised intelligence in efficiency improvement in the parcel vehicle industry.

Keywords: Decentralised freight intelligence, delivery vehicle drivers, parcel delivery efficiency, intelligent parcels, self-organising logistics, situational awareness

3 TABLE OF CONTENTS

1. INTRODUCTION ... 4

2. THEORETICAL BACKGROUND ... 6

2.1 The evolution of logistics planning systems ... 6

2.1.1 Traditional centralised planning ... 6

2.1.2 Traditional decentralised planning ... 7

2.1.3 Decentralised intelligent planning ... 7

2.1.4 The lack of empirical evidence of decentralised intelligence benefits ... 9

2.2 Parcel distribution efficiency ... 10

2.2.1 Key performance indicators ... 10

2.2.2 Situational awareness ... 11

2.2.3 The social aspect of the transition to decentralised intelligence ... 11

3. METHODOLOGY ... 13

3.1 Research setting ... 13

3.2 Experimental design ... 14

3.3 Data collection and analysis ... 16

4. FINDINGS ... 17

4.1 The effect of external information on search time ... 17

4.1.1 Segments policy ... 17

4.1.2 LIFO policy ... 19

4.2 Evaluating efficiency gains ... 20

4.2.1 Performance improvement ... 20

4.2.2 The impact of increased search time ... 21

4.2.3 The needed number of vehicles ... 23

4.3 Accuracy of external information ... 24

5. DISCUSSION AND CONCLUSIONS ... 26

5.1 Theoretical implications ... 26

5.2 Practical implications ... 27

5.3 Limitations and future direction ... 28

5.4 Conclusion ... 30

6. REFERENCES ... 31

7. APPENDICES ... 34

Appendix A: Layout of the segments in the vehicles ... 34

Appendix B: Overview of participating individuals in the experiment ... 34

Appendix C: Overview of search time data ... 35

Appendix D: Determination of volume capacity ... 41

4 1. INTRODUCTION

In 2018, every second 2,300 parcels were shipped globally and the retail e-commerce industry showed an annual growth of 23.3 percent (Bowes, 2018; Statista, 2019). This growth and the increasing pressure for delivery speed, reliability, and environmental focus (ITP, 2018) intensify the complexity of accurate and efficient handling and planning of parcel deliveries. In order to cope with this pressure and complexity, the development of novel planning concepts regarding the efficiency of the parcel vehicle industry seems crucial. These concepts are required to be both flexible and robust to respond quickly to changing circumstances and high delivery requirements and to avoid system breakdowns (Jedermann &

Lang, 2008; Kim, Sting, & Loch, 2014; Scholz-Reiter, Windt, & Freitag, 2004). As traditional planning systems are not sufficiently flexible and robust to process such a complex and large amount of information (Arendt, Klein, & Barwig, 2016; Costa et al., 2016; Moyaux, Chaib-Draa, & D’Amours, 2006), the need for a new system has emerged: a decentralised intelligent system.

Decentralised intelligent freight is cargo that is capable of making decisions autonomously (Sternberg & Andersson, 2014). Such intelligent goods are called agents and are described as anything that perceives its environment through sensors and provides a reaction on that environment through actuators (Russell & Norvig, 2016). Examples of these agents are components, products, pallets, parcels, trucks and containers (Scholz-Reiter et al., 2004). These agents are per definition intelligent as they are connected to many information technology (IT) services and as they have self, context, and location awareness. Through these new IT services, all agents in the entire network become able to communicate and make decisions based on information provided by relevant other agents (McFarlane, Sarma, Chirn, Wong, & Ashton, 2003). Practically, this implies that the agents contact others, such as vehicle drivers, to provide them with information about themselves and their decisions. For example, a parcel knows that its delivery vehicle is at a certain address of delivery, so the parcel lights up to communicate to the driver that he wants to be delivered. This communication can be received by the driver who uses, for example, augmented reality glasses.

There is an ongoing debate on the enhancement of decentralised intelligence in freight logistics.

Several benefits have been formulated in conceptual studies (Huschebeck, Piers, Mans, Schygulla, &

Wild, 2009; Sternberg, Hagen, Paganelli, & Lumsden, 2010), such as reduced driven kilometres and CO2-reduction. Furthermore, decentralized intelligence is expected to increase the system’s ability to quickly respond to changes. Besides customer service improvement, this may also lead to a decrease in harm to drivers as it alerts them for dangerous situations. However, empirical evidence of improved efficiency is not found yet (Sternberg & Andersson, 2014) and it is therefore not clear whether such benefits or other advantages would emerge if decentralised intelligence is used in freight logistics in practice.

The disadvantages of decentralised intelligence are mainly related to the implementation phase.

They are described as the massive implementation costs (Huschebeck et al., 2009), organisational complexity within the automation transition to decentralised intelligence (Zygmunt, 2010) and the

5 current barriers for multi-organisational adoption (Sternberg & Andersson, 2014). Critics state that the investment in decentralised intelligence may only be beneficial in settings which are too difficult for traditional planning systems to process (Jedermann & Lang, 2008). This could be the case when a high amount of local information, such as real-time data of perishable goods, needs to be collected. Another example is a setting in which confidential data, to which some human planners have no access, needs to be processed. Only in those settings, decentralised intelligence is claimed to be an option. Thus, literature remains inconclusive with regard to the benefits of enhancing decentralised intelligence in freight logistics.

Considering the limited empirical literature available, this thesis will take a first step in empirically investigating the effects of decentralised intelligence on efficiency in the parcel delivery industry. One aspect of decentralised planning systems is that local information is processed and used for decision-making. For example, a parcel has a cheap computer-on-a-chip with low power requirements, sensors, and wireless connectivity (Quak, Kempen, & Hopman, 2018). The parcel travels with DHL and becomes aware that it is the only parcel with destination Marken. When it notices that a UPS vehicle unexpectedly changes its route to Marken, the parcel decides to be transloaded to the UPS vehicle at a micro-hub, as that appears to be more efficient. As a result, the vehicle drivers are suddenly confronted with an unexpected change and lose therefore the awareness of what is happening inside the back of their vehicles. So, the responsive character of decentralised intelligence may reduce their situational awareness; the ability to understand the situation and react immediately on real-time events (Costa et al., 2016). This thesis’ main contribution regards therefore to understand the effect of the drivers’ loss of this situational awareness. The research question addressed in this thesis is as follows:

What is the effect of the loss of situational awareness of parcel vehicle drivers and how can this potential loss be encountered?

We conduct a field experiment among parcel delivery vehicle drivers during their tours. We empirically consider the effect of situational awareness loss at the addresses of delivery. Moreover, we investigate how we can diminish the potential negative effects of this loss, by testing whether external information can take over the role of the drivers’ situational awareness. We do so by providing drivers with information regarding the location of the parcels in their vehicles and measure whether it supports them to find the parcels at the addresses of delivery.

The remainder of this thesis is organised as follows. Section 2, the theoretical background, describes the existing knowledge of the need and constraints associated with decentralised intelligence and parcel vehicle efficiency improvement. The methodology section explains the experiment which examines the influence of additional information on efficiency. The fourth section contains the findings related to the experiment, and the final section of this thesis contains the discussion and conclusions.

6 2. THEORETICAL BACKGROUND

2.1 The evolution of logistics planning systems

With the ongoing changing process towards more dynamic and complex logistics, the structure of planning systems has evolved. A technology shift towards intelligent agents makes this evolution possible. This section provides an overview of the evolution to the decentralised intelligence planning structure and explains why literature still lacks empirical evidence regarding the benefits of decentralised intelligence.

2.1.1 Traditional centralised planning

In a traditional central planning system, the distribution of parcels is planned and scheduled from a central location, which processes information related to all actors in the distribution system (e.g.

parcels, vendors, customers, trucks, and shippers). Figure 2.1 provides an architecture of this system.

The parcels which flow through the system are controlled from this central location on top of a hierarchical planning structure (Scholz-Reiter et al., 2004). All information is directed to and from the central planner in the system, while the separate items do not communicate with each other. Such a system is comparable with the hierarchical organisational structure of the military (Moyaux et al., 2006).

However, several reasons prevent a centralised distribution system from operating efficiently (Fischer, Müller, & Pischel, 1996; Huhns & Stephens, 1999). The number of shipments and the market developments increase (Meyer, Kopfer, Kok, & Schutten, 2011), and as a result, so do the degrees of complexity and dynamics. The local information needed to plan the distribution can therefore be widely geographically distributed, may have a high degree of heterogeneity, and may be large in size.

Consequently, as the planner is directing from a central location, some local data are inaccessible to him or her, resulting in incomplete information and a limited viewpoint (Jennings, Sycara, & Wooldridge, 1998). So, a centralised planning system is not sufficiently flexible to cope with the growing complexity and dynamics (Arendt et al., 2016; Costa et al., 2016).

Figure 2.1 Traditional centralised distribution architecture

7 2.1.2 Traditional decentralised planning

The difficulty to handle the complex and dynamic market developments has led to the shift from centralised to decentralised distribution systems (Scholz-Reiter et al., 2004). These decentralised distribution systems are characterised by multiple planners who all take control over a certain part of the system (see Figure 2.2). Due to its modular design, the distribution in a decentralised system is easier to understand for (IT-steered) planners and therefore more flexible (Huhns & Stephens, 1999). Each computation requires less information than in the centralised approach, implying that the software is easier to write and to debug in cases of changes in the system (Moyaux et al., 2006).

Nevertheless, this system carries a three significant disadvantages. Firstly, a global system control is missing; local information belonging to a certain district is isolated and not interconnected with other districts. Accordingly, parcels travelling through several districts are likely to take inefficient routes and be delayed (Van Dyke, 1994). Secondly, a decentralised distribution system can create local optima within districts, but parcels often travel through multiple districts and the different decentralised routing plans are not able to synchronise optimally as a whole (Moyaux et al., 2006). For example, a parcel with a radio-frequency identification (RFID) tag leaves one district and enters another. The parcel first needs to travel to a main location within the new district to be scanned, before it can be planned for distribution within the new district. This may lead to large detours and long waiting times. Finally, predictions for each planner can only be made at an aggregate level, instead of an individual level (Huhns

& Stephens, 1999). Although the market is shifting towards a more dynamic state, this means that individual customer requirements are difficult to be taken into account.

Figure 2.2 Traditional decentralised distribution architecture

2.1.3 Decentralised intelligent planning

Since centralised distribution systems cannot handle the growth and complexity and decentralised distribution systems cannot operate efficiently, research has proposed a novel concept:

decentralised intelligent planning (Scholz-Reiter et al., 2004). This is a self-organising system without one or more central planners. Instead, the planning processes are performed by all participating agents, including parcels. Such parcels are non-isolated goods that process and communicate information by

8 means of storing data concerning themselves and by making decisions with a local scope (McFarlane et al., 2003; Sternberg & Andersson, 2014). Their intelligence is decentralised because they only collect information which is relevant for their own decision-making. Such information and local-scope decisions could be real-time traffic data on the potential routes between their origin and destination, and the choice for a certain vehicle for a pick-up. Decentralised intelligence is different from existing systems featured by ICT solutions such as RFID. Although RFID-tagged parcels have unique identities which can be read by an information system, they do not have the ability to gather and process data from other resources. Such parcels are therefore not sufficiently intelligent to make optimal decisions (Sternberg & Andersson, 2014). So, decentralised freight intelligence goes beyond these kind of existing ICT solutions.

Figure 2.3 depicts the architecture of an intelligent decentralised logistics platform, which consists of a number of actors (Gallay, Korpela, Tapio, & Nurminen, 2017). Vendors, customers, shippers, and parcels provide the initial information, which flows bottom-up through the process.

Containers, vehicles and parcels are receivers of this information in terms of real-time data, positions of the actors, and other sensory information. Shippers and customers receive the tracking information.

Several IT services manage these data flows through the network. Industrial Data Space technology secures the data exchange, Internet-of-Things technology gathers and communicates real-time data over the network, and a meta-connector processes all information and uses that to search, optimise, and manage the routes and processes through several transportation modes and companies.

Figure 2.3 Proposed decentralised intelligent distribution architecture

Several potential advantages of such an intelligent decentralised logistics platform have been identified (Gallay et al., 2017). Firstly, the Internet-of-Things enables real-time estimation of time of arrival, implying that carriers, logistics service providers, and customers can adjust their planning on

9 this estimation. These adjustments decrease lateness and thus a better score on on-time delivery can be achieved. Secondly, the availability of real-time data allows parcels to provide information regarding their location and condition. Sensors are able to track data, which connect the parcel to the network.

This may lead to practical improvements such as efficient transloading, routing, tracking, security and technological integration (e.g. Gallay et al., 2017). Finally, the platform allows integration of current transport management systems and mobile applications, which increase customer service.

As there is no centralised or decentralised planner, decentralised intelligence is depending on rules and protocols, which assure decentralised control (Costa et al., 2016; McFarlane et al., 2003;

Rekersbrink, Makuschewitz, & Scholz-Reiter, 2009). That means that an agent can be compared to an actor in the traffic such as a car driver or a cyclist, whereas the decentralised intelligent architecture can be compared to the road infrastructure (Bonomi, Milito, Zhu, & Addepalli, 2012). Cars and cyclist all have their own destination and pick their own route. Even though they receive information from navigation devices, they make their own decisions. The decision for the route is motivated by individual wishes, such as the shortest route, the fastest route, the most sustainable route, or the most touristic route. Driving from A to B, the cars and cyclists have to follow the traffic rules. Correspondingly, in a decentralised intelligent distribution system, agents make their own decisions concerning their most convenient route, based on predetermined algorithmic settings such as time-efficiency or cost- efficiency.

2.1.4 The lack of empirical evidence of decentralised intelligence benefits

Although decentralised intelligence in freight logistics is a relatively new literature field, a broad range of research has been conducted. Several authors have investigated how decentralised intelligence is technically conditioned (Gallay et al., 2017; Pan et al., 2016). For example, sensors and actuators have to function inside a decisional loop in order to enhance autonomous acting. Other research has been performed regarding the benefits and disadvantages of decentralised intelligence (Huschebeck et al., 2009; Janssen, 2004; Jedermann & Lang, 2008; Pan et al., 2016; Sternberg et al., 2010). Sustainability seems to be the most beneficial factor from the efficiency that decentralised intelligence is expected to bring, whereas the implementation phase is expected to be the largest barrier for the application of decentralised freight intelligence. Even practical approaches have been considered (Quak et al., 2018;

Scholz-Reiter et al., 2004), including the support that decentralised intelligence may offer to parcels for self-organisation in depots.

Nevertheless, we stress three critical shortcomings in literature. Firstly, none of the current known scientific papers has an empirical approach. It is therefore unclear how decentralised intelligence can be adopted by the logistics industry (Sternberg & Andersson, 2014). The lack of available quantitative data is likely to be the reason for its absence in literature (Holmström, Prockl, & Sternberg, 2014). Secondly, little significant support has so far been found in favour of efficiency improvement by decentralised intelligence in logistics (Sternberg & Andersson, 2014; Sternberg & Norrman, 2017).

Thirdly, few authors have taken into account the social impact of decentralised intelligence, which

10 regards that the intelligence of actors take over some of the decision-making abilities of humans. The implementation of such a technology may be considered as a radical disruption (Hendersson & Clark, 1990), which may have a large impact on both the aggregate level of society and on the individual level of humans.

2.2 Parcel distribution efficiency

The empirical approach of this thesis focusses on how local information can be provided in a good manner for the purpose of improving performance. By doing so, we follow the advice of Sternberg and Andersson (2014) to investigate how IT services could be managed on the individual level, implying that they are not depending on other agents. This section takes off with an overview of what is known in literature regarding performance in the specific field of parcel distribution. Next, we present the understanding of situational awareness of humans and how that may be taken over by agents when decentralised intelligence is implemented. That leads to the final sub section related to the social impact of decentralised intelligence.

2.2.1 Key performance indicators

The enhancement of decentralised freight logistics is proposed to improve freight transport efficiency (Lumsden & Stefansson, 2007). Using key performance indicators (KPI) is important to evaluate these improvements (Davidsson, Clemedtson, Mbiydzenyuy, & Persson, 2012). Four KPIs in the parcel industry are minimisation of the number of vehicles, the total travel distance, the total time, and waiting time at customers (Moura & Oliveira, 2009). Besides that, little attention in literature is paid to the performance of parcel vehicles.

Nevertheless, a lot of research is carried out in the parallel field of heavy goods vehicles operating in international terminals that handle containers (Davidsson et al., 2012; Fazili, Venkatadri, Cyrus, & Tajbakhsh, 2017; Koç, Bektaş, Jabali, & Laporte, 2016; Mbiydzenyuy, 2013). Examples of general KPIs related to vehicle performance on the road are fuel costs, distance-based costs, time-based costs, transport administration, accidents, infrastructure maintenance costs, noise, building of new infrastructure, and cost of missing and delayed goods (Davidsson et al., 2012). The new concept of intelligent freight generated new perspectives concerning performance indicators. Additional road performance indicators regard therefore the reduction of CO2-emissions, learning capabilities of agents, and large-scale opportunities as positive effects of decentralised distribution systems (Costa et al., 2016).

Other KPIs relate to performance the handling in the terminals, depots, or at customers. A major part handling work is unloading of the vehicles. Unloading is a task that takes place in the last mile. As forty percent of all logistics activities take place in the last mile, efficiency improvements in the last mile can make a large impact (Roumboutsos, Kapros, & Vanelslander, 2014). Unloading in the last mile requires more time-efficiency, because authorities (e.g. municipalities) set stricter rules that plead for shorter parking times for parcel delivery vehicles. This is caused by the trend of improving green city

11 logistics that has emerged among both private and public organisations (De Marco, Mangano, &

Zenezini, 2018).

Yet, performing well on these KPIs is constrained by several factors, accounting for both traditional as newly emerged KPIs. Acceptance of inventions, high turnover of staff, unavailable customers, traffic incidents, strict time frames, and ad-hoc orders (a last-minute change in a customer’s time-frame requirement) are significant constraining determinants of parcel delivery optimisation (Menge & Hebes, 2011).

2.2.2 Situational awareness

Agents are designed in a manner that they desire to achieve objectives such as delivery reliability and cost-efficiency (Jennings et al., 1998), because they have situational awareness (Costa et al., 2016).

The agents are able to achieve these objectives by making certain decisions (Baykasoglu & Kaplanoglu, 2011). They firstly sense real-time events, after which they interpret the situation of interest, and then they consider possible reactions in order to execute a plan or solution (Costa et al., 2016). For example, an intelligent parcel communicates to the system that it needs to travel from A to B. To solve this problem, the parcel waits for several offers from shippers and selects the best offer based on its individual selection criteria (e.g. minimisation of emissions). However, a change in the plans of the shipper might change the contract between the parcel and the shipper.

Agents are both reactive, because they are able to react to real-time changing situations, and social, because they communicate with humans, such as drivers (Naji, Etzkorn, & Wells, 2004). On the one hand, a last-minute change of plans from an agent needs to be clearly communicated to the involved humans, because the humans lose their situational awareness as the consequence of such an unexpected change. On the other hand, real-time events may be caused by the decisions of humans. These decisions can be considered as a data inputs processed by the agent. Collecting and processing these data inputs support the agents to become more aware of their situation.

It is unclear whether the agents’ gain of situational awareness outweighs the drivers’ loss of situational awareness in terms of efficiency. Agents are highly situational aware, because they are able to base their decisions on a large amount of information and to make use of services for real-time rerouting (Costa et al., 2016). The agents make decisions which are outside the control of the drivers.

Concluding, although there are several arguments in favour of the performance resulting from agents’

situational awareness, little is known regarding the drivers’ reactions and the behaviour of that loss of control. Their reactions may lead to changes in information which is sensed by agents as real-time events, which finally may influence the decisions that the agents take.

2.2.3 The social aspect of the transition to decentralised intelligence

One of the challenges in the transition to an IT-steered distribution system is the acceptance of operators and vehicle drivers (Menge & Hebes, 2011). If agents make decisions on their own, drivers have less authority and decision-making abilities. A study focussed on minimising greenhouse

12 emissions showed that the success of an innovation or invention depends on the acceptance of drivers (Menge & Hebes, 2011). Especially with freelance drivers, this acceptance is hard to gain, as they are mostly focussed on maximising their own benefits. This implies that the implementation of a new system should be beneficial for the drivers as well.

Fortunately, drivers are likely to be the stakeholders who benefit most from the implementation of an intelligent transport system (Mbiydzenyuy, 2013). If information is communicated clearly, such as in the proposed decentralised intelligence system, the drivers’ perceived safety increases by lowering stress and uncertainty. Furthermore, decentralised intelligence reduces the number of accidents and reduces the time needed to find a place to park the vehicle. However, most of these benefits are long- term visions. Resistance to initial implementation is therefore still high. The limited acceptance of inventions is one of the major constraints in parcel service optimisation (Menge & Hebes, 2011).

Therefore, it is interesting to find out what the loss of situational awareness has on parcel vehicle drivers.

13 3. METHODOLOGY

3.1 Research setting

In a decentralised intelligence planning system, agents collect and process information and make decisions which are currently made by humans. We expect that this implies that humans’ situational awareness decreases. In this thesis we explore the effect of the loss of parcel vehicle drivers’ situational awareness and we consider how we can encounter this potential loss. To that end, we conduct a field experiment, supported by observations and interviews.

The experiment is performed at a Dutch parcel delivery service company. The company has several depots in the Netherlands. The depot in which the experiment is performed is temporarily in use¹, due to both the limited capacity of a large depot located nearby and the fast growth of the company.

Most of the sorting work in this depot is performed manually at a conveyor belt. The area which is served from this tour is divided into approximately forty districts of which each is served by the same driver every day, regardless of the number of parcels ordered in that district. Two of the freelance drivers, whom the company works with, have volunteered to participate in this research. Practical implications have limited this number of drivers in this experiment. The intensive character of the experiment was quite invasive on the daily working conditions. We are therefore content with the six full days during which the drivers and the shipping organisation were willing to participate.

The operational processes at the focal company are performed as follows. During the night, several inbound batches arrive at the depot with parcels ready to be distributed to the customers the subsequent day. After arrival, these parcels are sorted by the depot operators into the forty tour sectors on the floor. Drivers arrive at the depot between 07.00 and 07.30 A.M. On most days, the sorting process is already finished at this time, but during busy days the drivers need to wait for the sorting process to finish. As soon as all parcels are sorted in the tour sectors, drivers scan the parcels for their own tour.

During scanning they resort the parcels in their tour sector on the floor, so that parcels are sorted together which belong to addresses located near each other. They do this based on their own knowledge and recognition of the addresses. Next, the planning software calculates the route and takes into account factors such as opening times and time tables regarding the allowance of driving in city centres. Most of the drivers adapt this route according to their preferences. The drivers often have local information, for example, they take the back door instead of the front door at delivery or they change the route, because they know their vehicles are too large for a certain bridge. Finally, the drivers load their own vehicles with the grouped parcels together; in a large number of small segments. They do this in same manner every day. This helps them to remember where they have located the parcels and it increases their situational awareness, which makes it easier to quickly find the parcels at the addresses of delivery.

1 Planned for a period of three years

14 3.2 Experimental design

The goal of the experiment is to explore the consequences of reduced situational awareness of the drivers, which result in more time on the road, but longer search times. Time currently spent on the loading process² is valuable time which the drivers currently cannot spend on making their tours.

Consequently, their time on the road could be longer if their vehicles are loaded by someone (i.e. a depot operator) or something (i.e. a robot) else. We reflect on how the search time of drivers is affected if the loading process is assigned to depot operators.

Firstly, we test how external information can reduce the impact of situational awareness loss.

We provide different levels of information by using two loading policies. In the experiment, parcels are being loaded in the vehicle according the segments and last-in-first-out (LIFO) policies. These policies are depicted in a schematic overview in Figure 3.1. Loading in segments implies that the parcels are sorted into six segments within the tour sector at the depot. The six segments are based on the first four or five digits of the postal codes and contain one or several codes (see Appendix A). The drivers are provided with the external information that the parcel is located in a certain segment. However, they have no information regarding the parcel location within the segment (i.e. front, back, upper or lower part).

Figure 3.1 Schematic two-dimensional vehicle top view of loading policies

In the LIFO policy, the parcels are loaded exactly according to stop number, meaning in the right order in which they are planned to be unloaded. The parcel should be directly accessible for anyone looking for that parcel. In this experiment, LIFO-loaded parcels are also provided with a written stop number on the parcel. This indicates a higher accuracy level of external information than for segments- loaded parcels. The drivers have more information from the LIFO policy, because they know that the needed parcel is always the first one that the they see after opening the vehicle, it is located near the previous parcel, and the driver can quickly check whether it is the right parcel by looking at the written

2 The process of scanning, planning optimisation, and loading

15 stop number on the parcel. The link between the loading policies and external information is represented in Figure 3.2. They are compared to the situation in which the drivers scan and load their own parcels (i.e. self-loading). This is how the drivers are used to operate, meaning they have a lot of situational awareness and they have little need for external information, as they remember where they have located the parcels during the loading process.

Figure 3.2 Approximate representation of decentralised intelligence per loading policy

Secondly, we investigate what level of accuracy of external information is necessary to achieve reasonable efficiency. Regarding the software-proposed route planning, we make a distinction between with and without adaptions by the driver. Route planning with adaptions implies that drivers may adapt the proposed route based on their own preferences, which is how they are used to work. In the settings without route adaptions by the drivers, we restrict drivers from adapting the route according to their preferences and therefore oblige them to drive the route as proposed by the software. By means of interviews we determine the drivers’ experiences related to the software. A negative experience with the software-proposed route is expected to imply that the software does not process accurate information, meaning that drivers experience that the proposed route does not lead them along the most efficient route.

The experiment is conducted in two scenarios. The first scenario tests the segments policy and the routing policy by which the drivers are allowed to adapt the route. The second scenario tests the LIFO policy and the software-proposed route without adaptions by the drivers. The loading policies affect the handling in the depot and the handling within the vehicles at the addresses of delivery, whereas the routing policies have effect on the total time spent on the route and the order in which the customers are being served. Therefore, the loading and routing policies were not expected to influence each other during the experiment and were tested simultaneously in both scenarios.

Both scenarios are split into two sub scenarios. Table 3.1 provides an overview of the scenarios and their experiment components. Base-scenarios contain the self-loading policy and represent the baseline measurements. The intelligent-scenarios are the scenarios in which depot operators perform the loading process. The local information is necessarily provided, because the drivers have not seen the parcels before and are dependent on the provided information to find the parcels. The base-scenarios are tested for one day, because self-loading is the usual way of operating. No errors were made because the drivers are experienced in scanning and loading their own parcels. The intelligent-scenarios are tested for two days, to take into account the learning effect of the depot operators and drivers; a large difference in results between both days, may indicate that the drivers and operators made mistakes, because they needed to get used to the new working methods.

Low High

Self-

loading Segments-

loading LIFO- loading Accuracy level of provided information

16 Table 3.1 Experiment components

Scenario Loading policy Routing policy Performer of loading process

1 Base-segments Segments Driver Driver

Intelligent-segments Segments Depot operator Depot operator

2 Base-LIFO LIFO Driver Driver

Intelligent-LIFO LIFO Depot operator Depot operator

3.3 Data collection and analysis

During each scenario, the search time is measured by the researcher in one vehicle and an assistant-researcher in the other vehicle. The time is measured with a stopwatch from the moment the driver opens the backdoor or side door until he or she found the parcel. A second, third etc. parcel is measured from the moment the driver found the previous parcel until he or she found the searched parcel. Pick-up stops were neglected, because no parcels needed to be searched in the vehicles during pick-ups. In the analysis of the effect of the search time, averages were calculated in order to diminish the effect of differences between days and drivers. In addition, frequency graphs and chronological graphs were made in order to find patterns in the data and consider the drivers’ working methods.

Secondary data were provided by the focal company. These contained real time data for their share of the Dutch parcel delivery market for the period from 10 August 2018 until 16 August 2018.

The used data contained drivers’ productivity, scan amounts, stop time and tour length. Data were analysed and integrated with the collected data in Microsoft Excel® 2016. A driver’s working day was analysed and broken down into several processes. Area time is one of these processes, which consists of road time and stop time, of which the latter consists of search time and non-search time. The search time measurements for the different loading policies were applied to the work day processes in order to see the effect on the daily number of parcels to deliver. Thereafter, a correction for volume capacity and flexibility levels of volume capacity was executed.

Before and during the experiment, observations and interviews were held with drivers and operational managers (see Appendix B). Their answers were examined after the analysis of the findings, in order to examine the potential reasons behind the quantitative data. This qualitative part of the data collection had four goals. Firstly, we aimed at determining KPIs for the parcel delivery industry. Moving towards a logistics system with a higher level of decentralised intelligence can only be beneficial if it positively affects the KPIs. Secondly, we take into account to what extent drivers are affected by giving away their situational awareness and control. Thirdly, we consider the practical implications of assigning the loading process to depot operators. During the intelligent-segments scenarios, the loading process was assigned to inexperienced depot operators, whereas during scenario intelligent-LIFO, these tasks were done by experienced colleague drivers. In this way, we tested the effect of loading mistakes.

Finally, we analyse the differences, and the reasons behind these differences, between the routes proposed by the software and those adapted by the drivers. This indicates the needed level of accuracy of the information provided externally. This is relevant for the implementation of decentralised intelligence, as local intelligence will have to be transformed from drivers’ knowledge to parcel intelligence.

17 4. FINDINGS

This section provides a review of the findings from the experiment; the data analysis, the observations and the interviews. A complete overview of the findings can be found in Appendix C.

Firstly, the effect of the two loading policies on search time is discussed, including the patterns which show the working method of drivers to search for the parcel. We attempt to find out whether external information would help drivers to find the parcels in their vehicles. Subsequently, we discuss the analysis of the common KPIs in the parcel delivery industry. We also take into account time and volume capacity constraints with regard to achieving these KPIs. Although drivers’ search time may increase, we do not expect drivers to spend more time to their working day. Besides, even though assigning the morning process to depot operators might relieve more time for the drivers to deliver parcels, their vehicles do not always have enough space to carry all these parcels. Thereafter, we discuss how many vehicles we expect to need. Finally, this section discusses the accuracy level of information that drivers need to quickly find the parcels in their vehicles.

4.1 The effect of external information on search time 4.1.1 Segments policy

Finding the parcels in a segment when the drivers did not load their own vehicle (intelligent- segments scenarios), resulted in a longer search time compared to the base-segments scenario with self- loading (see Table 4.1). This is the consequence of the limited provided external information.

Table 4.1 Search time results segment scenario Scenario Average search time

per parcel (m:ss)

Base-segments 0:05

Intelligent-segments 0:16

Looking into more detail, significant differences can be found between the base-segments and the intelligent-segments scenarios. The results in Figure 4.1 are combined results from both participating drivers during all researched days, so that the difference in experience from drivers is inconsequential.

Under self-loading, drivers could find approximately 81% of the parcels immediately (i.e. < 5 seconds).

In the sub scenario in which the depot operators loaded the vehicles, drivers could find ca. 49% of the parcels immediately. Approximately 13% of the parcels (compared to 2% in the base-segments scenario) required more than 30 seconds to find.

18 Figure 4.1 Search time frequency diagram of segments scenarios

During their search for a parcel in the intelligent-segments scenario, the drivers had to search throughout the entire segment. However, because they were not aware of the visuals and the exact location of the parcels, they experienced difficulties to find the parcel within the segment. So, their working method was to reorganise the parcels every few stops. This is illustrated by the peaks shown in Figure 4.2. This graph represents one of the days of one of the drivers under the intelligent-segments scenario. Although the vehicle contained the relatively small number of 65 parcels, it was still not doable to find all parcels immediately. The search time was short for most of the stops shortly after the peaks.

That was the gain from the long search time during the peaks, in which the drivers collected more information from multiple parcels related to their visuals and location. Concluding, in the segments policy, the level of information from the parcels was not sufficiently high for the drivers to quickly find all the parcels.

Figure 4.2 Timeline example of intelligent-segments scenario

0% 10% 20% 30% 40% 50% 60% 70% 80% 90%

0:00 - 0:05 0:05 - 0:10 0:10 - 0:15 0:15 - 0:20 0:20 - 0:25 0:25 - 0:30

> 0:30

Percentage of parcels

Search time (m:ss)

Intelligent segment Base segment

0:00 0:17 0:35 0:52 1:09 1:26 1:44 2:01 2:18 2:36 2:53

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63 65

Search time (m:ss)

Parcel (chronological)

19 4.1.2 LIFO policy

As shown in Table 4.2, no difference is found in search time between self-loading and loading by depot operators following the LIFO policy. The level of external information is high during the intelligent-LIFO policy, because stop numbers are written on the parcels, and accurate location is known, which is next to the location of the previous parcel.

Table 4.2 Search time results LIFO scenario Scenario Average search time

per parcel (m:ss)

Base-LIFO 0:06

Intelligent-LIFO 0:06

The search times of the base-LIFO and intelligent-LIFO scenarios show similar patterns, regardless of whether the drivers or the depot operators loaded. This can be seen in Figure 4.3, which depicts the distribution of parcels over the different ranges of search time under the LIFO policy.

Interestingly, drivers could find relatively few parcels within five seconds in the base-LIFO scenario, compared to the base-segments scenario. This may be due to the fact that the loading policy that they are used to work with, is more similar to segments-loading. However, their average search time under self-loading is only one second more for LIFO than for segments. This difference can be neglected, because it seems insignificant.

Figure 4.3 Search time frequency diagram of LIFO scenarios

The timeline of one of the days during the intelligent-LIFO scenario (Figure 4.4 Timeline example of intelligent-LIFO scenarioFigure 4.4) teaches us that drivers barely need any time to reorganise the parcels. This implies that the provided external information is sufficiently accurate, meaning there was no need to get familiar with the parcels in the vehicles. Drivers could find a parcel quickly, because it was the first parcel that the driver noticed, it was located near the previous parcel, and the driver could verify the stop number by looking at the written number on the parcel.

0% 10% 20% 30% 40% 50% 60% 70% 80%

0:00 - 0:05 0:05 - 0:10 0:10 - 0:15 0:15 - 0:20 0:20 - 0:25 0:25 - 0:30

> 0:30

Percentage of parcels

Search time (m:ss)

Intelligent LIFO Base LIFO

20 Figure 4.4 Timeline example of intelligent-LIFO scenario

In conclusion, the intelligent-LIFO scenario provided similarly accurate information as the base- LIFO scenario. To compare, parcels loaded in segments provide such inaccurate information that drivers needed almost three times as long to find the parcels. Therefore, the segments policy represents a lower level of decentralised intelligence than the LIFO policy. A lower level of decentralised intelligence related to the parcels’ locations in the vehicles leads to a longer search time. That means that under a decentralised intelligence system, parcels should make accurate and logical decisions concerning which location in the vehicle they choose and provide clear communication to the drivers.

4.2 Evaluating efficiency gains 4.2.1 Performance improvement

Based on the interviews, the KPIs as described in Table 4.3 are considered important within the parcel delivery industry. Only the indicators related to the drivers are taken into account. Assigning the loading process to depot operators affects the drivers’ depot time positively, since the drivers do not spend any time at the depot in the morning. Besides, the stop time will be affected negatively, since the search time slightly increases. Furthermore, the number of driven kilometres decreases, since fewer vehicles are needed on the road. A large part of the driven kilometres is driven between the depots and the delivery areas. Although the number of kilometres in the delivery area grows as the number of deliveries per tour grows, the total number of driven kilometres still declines, since the number of rides between the depot and the delivery area decreases. In this thesis, we focus indirectly on decreasing the driven kilometres, by considerably decreasing the depot time of drivers, so they can spend more hours to their tours, and by slightly increasing the stop time (search time).

0:00 0:17 0:35 0:52 1:09 1:26 1:44 2:01 2:18 2:36 2:53

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94 97 100 103 106 109 112

Search time (mm:ss)

Parcel (chronological)

21 Table 4.3 Key performance indicators in the parcel delivery

industry

KPI Description

Depot time Time drivers spend at the depot during the morning process

Driven kilometres Total distance drivers drive Productivity Percentage of stops that have been

fulfilled successfully (hit rate) Stop time Time drivers spend at the address of

delivery

Waiting time Time drivers need to wait before they are allowed to arrive at a pick-up address

4.2.2 The impact of increased search time

This section discusses the effect of different search time means on the overall efficiency of the parcel delivery vehicle. We take into account the time and volume capacity constraints to estimate the potential performance improvements in terms of needed vehicles. Assigning the loading process to depot operators increases the time required for the drivers to deliver parcels. Drivers start their working day later and gain efficiency because they have more time to deliver parcels. However, drivers lose efficiency because their search time increases. Because it is unrealistic for drivers to have longer working days, we have to take the time capacity constraint into consideration.

A driver’s working day can be divided into seven processes. Figure 4.5 provides an overview of the average working day including the percentage of time each process takes. The loading process is the process of scanning and loading the parcels and adapting the software-proposed route. The way up is the time needed to drive from the depot to the first address of delivery/pick-up. The area process is the time drivers spend in the area for delivery and pick-ups. Breaks is the time drivers spend on a break.

Most of the drivers do not schedule time for breaks, but they have their lunch while driving instead.

Waiting time involves the time drivers need to wait before they are allowed to arrive at a pick-up address.

The way down is the time needed to drive from the final address of delivery/pick-up to the depot. The unloading process is the process of unloading the picked-up parcels at the depot.

Figure 4.5 Drivers' working day processes

The average area process consists of stop time, time spent at the addresses of delivery, and road time, time spent on the road, i.e. time not spent at the address of delivery. As shown in Figure 4.6, the

16,85%

5,62%

58,43%

5,62%

3,37%6,74% 3,37%

Loading process Way up Area process Break Waiting time Way down Unloading process

22 largest part of the area time is stop time (57.29%), implying that improvements in this field can gain a lot of time-efficiency. Figure 4.7 shows how the stop time is divided between search time and non- search time for each researched scenario. The calculations are based on the averages of the several days and drivers for each scenario.

Figure 4.6 Division of area process time

Because the drivers are able to leave their vehicles at the depot during the night, possibilities are created to assign the loading process to depot operators. This implies that drivers can start their working day when their vehicles are already loaded and finish their working day at a later time of the day, without spending more hours working. As drivers currently start early in the morning (ca. 07.00 A.M.), they will still have normal working hours if they start working later. Summing up the morning process (16.85%) and the area process (58.43%) means that drivers can spend 75.28% of their day to their area. Consequently, they can spend more time on delivering parcels.

However, the number of extra parcels to deliver is not equal to the morning process (16.85%), implying that the reduction of needed vehicles is also smaller. To clarify, as drivers do not load their vehicles anymore, their search time increases (Figure 4.7). As a result, their stop time also increases, whereas their road time does not increase (Figure 4.6). So, the average time needed to deliver one parcel slightly increases. However, a drivers’ working day is restricted by time capacity. Since we cannot expect driver to have a longer working day, the area time is not stretchable. Therefore, an increased stop time leads to a decrease in road time, meaning that fewer customers can be visited. Taking this into account, the number of needed vehicles based on the eliminated morning process for drivers and the

57,29%

42,71%

Stop time Road time

Figure 4.7 Division of stop time 5,23%

94,77%

Base

15,11%

84,89%

Segments

Search time Non-search time

5,55%

94,45%

LIFO

23 augmented search time can be found in Table 4.4. This is different for both loading policies, since they result in different search times.

Table 4.4 Vehicles needed only based on time capacity as percentage of currently used vehicles

Segments LIFO Minimum needed vehicles 86.82% 77.21%

Although drivers can now deliver more parcels on average, not all vehicles have sufficient volume capacity. The maximum physical volume capacity of vehicles is 160 parcels (see Appendix D).

Theoretically, under dynamic planning, each vehicle should be able to deliver 160 parcels. The current planning system at the focal company suggests that each tour covers the same geographical area every day, regardless of the number of orders made by customers in that area for that day. If we include dynamic planning, tour planning is not based on geographical areas, but on for example equal number of stops, equal planned kilometres, or equal expected tour time. In this thesis, we assume that dynamic planning implies an equal number of parcels per tour. However, there is a trade-off between flexibility level and needed vehicles on the road. If each vehicle is planned to deliver 160 parcels, all flexibility is eliminated. Examples of situations in which flexibility is needed are that tours contains more large vehicles than averagely, or that pick-ups are planned at the beginning of the day, when not enough space is created in the vehicle yet.

4.2.3 The needed number of vehicles

To summarise, the following factors are taken into account for the calculation of the needed number of vehicles. Firstly, the added morning time to the area time led to an increase from 58.4% to 75.4% of time spent during the area time. Secondly, the time capacity was considered. The different loading policies led to different results in search time. The longer the search time per parcel, the longer the stop time, the fewer parcels to be delivered within the area time. Finally, the volume capacity per vehicle under dynamic planning was taken into consideration. Significant improvements can be gained, since the loss in search time is only a slight loss compared to the gain of the morning time in the depot.

Besides, only a few vehicles are constrained by volume capacity (Appendix D).

Table 4.5 provides the indication for the expected performance improvement based on the data we collected. The best case is considered to be the LIFO policy with the maximum allowance of 150 parcels per tour. In this case, the level of flexibility is sufficiently high, so that there is always space for larger parcels or larger orders. The need for only 80.39% of the vehicles compared to the current average has positive implications. Fewer vehicles on the road leads to lower emissions, fewer traffic jams, improved city logistics, and lower costs (Cagliano, De Marco, Mangano, & Zenezini, 2017; Costa et al., 2016; Davidsson et al., 2012). The costs needed for depot operators to load the vehicles are not included.

However, expectations are that these financial costs (e.g. wages, training, overhead) are lower than the savings gained from the decrease in the number of vehicles on the road under dynamic planning.