XIAO Haining，WU Xing，ZOU Ting，ZHAI Jingjing

1.School of Mechanical Engineering，Yancheng Institute of Technology，Yancheng 224051，P.R.China；

2.College of Mechanical and Electrical Engineering，Nanjing University of Aeronautics and Astronautics，Nanjing 210016，P.R.China；

3.Department of Mechanical Engineering，Memorial University of Newfoundland，St.John’s A1B 3X5，Canada

Abstract: In recent years，multiple-load automatic guided vehicle（AGV）is increasingly used in the logistics transportation fields，owing to the advantages of smaller fleet size and fewer occurrences of traffic congestion.However，one main challenge lies in the deadlock-avoidance for the dispatching process of a multiple-load AGV system. To prevent the system from falling into a deadlock，a strategy of keeping the number of jobs in the system（NJIS）at a low level is adopted in most existing literatures. It is noteworthy that a low-level NJIS will make the processing machine easier to be starved，thereby reducing the system efficiency unavoidably. The motivation of the paper is to develop a deadlock-avoidance dispatching method for a multiple-load AGV system operating at a high NJIS level. Firstly，the deadlock-avoidance dispatching method is devised by incorporating a deadlock-avoidance strategy into a dispatching procedure that contains four sub-problems. In this strategy，critical tasks are recognized according to the status of workstation buffers，and then temporarily forbidden to avoid potential deadlocks. Secondly，three multiattribute dispatching rules are designed for system efficiency，where both the traveling distance and the buffer status are taken into account. Finally，a simulation system is developed to evaluate the performance of the proposed deadlock-avoidance strategy and dispatching rules at different NJIS levels. The experimental results demonstrate that our deadlock-avoidance dispatching method can improve the system efficiency at a high NJIS level and the adaptability to various system settings，while still avoiding potential deadlocks.

Key words：automatic guided vehicle（AGV）dispatching；deadlock avoidance；multiple-load AGV system；critical tasks；multi-attribute rules

0 Introduction

Transportation system based on AGVs has been adopted as efficient，flexible and agile material handling systems in different environments，e.g.，manufacturing plants，warehouses，distribution cen?ters，and transshipment terminals. The design and control processes of an AGV system involve quite a few issues，i.e.，guide path designing［1］，vehicle dis?patching［2］，path/route planning［3］，and traffic man?agement［4］.

The AGV system dispatching problem has been a hot area of academic research since the mid-1980s. Various dispatching methods have been de?veloped. These dispatching methods can be divided into two types：offline and online. For the offline type，the task and vehicle data for the dispatching problem are known in advance. The entire dispatch?ing scheme can be optimized in advance. For exam?ple，Umar et al.［5］investigated the simultaneous scheduling of machines and dispatching AGVs in a flexible manufacturing system（FMS）environment.A mixed-integer nonlinear programming model was formulated firstly，and a hybrid multi-objective ge?netic algorithm was then proposed to optimize the makespan，AGV travel time，and penalty cost. Ku?mar et al.［6］addressed simultaneous scheduling of both machines and AGVs with alternative machines for the makespan minimizing objective. However，in practice，the manufacturing environments are usu?ally stochastic. A small change in job arrival time or a failure of an AGV may destroy the whole dispatch?ing plan. In this sense，online dispatching methods based on heuristic dispatching rules have more prac?tical robustness. Maniya et al.［7］proposed a multi-at?tribute dispatching method based on the AHP/MGRA technique. Xiao et al.［8］proposed a real-time deadlock-free multi-attribute dispatching method（RMDM）using three attributes：empty traveling distance，remaining space status of both input and output buffers. Nevertheless，most of these methods are only appropriate for a single-load AGV system.

Recently，several studies have showed that the utilization of multiple-load AGVs has many advan?tages，such as less traffic congestion［9］，smaller fleet size［10-11］，and increased system throughput［12］.Hence，the application of multiple-load AGVs is be?coming increasingly popular in modern factories，which also encourages the theoretical research activi?ties in academic communities.

Ho et al.［13］proposed a dispatching procedure for a multiple-load AGVs. In this procedure，task determination rules［14］， delivery dispatching rules［14］，pickup dispatching rules［13，15］and job selec?tion rules［15］were designed for four problems，which were evaluated via computer simulation. The simu?lation results indicated that job selection rules and pickup dispatching rules interact with each other［15］.Then，a multiple-attribute method［16］was proposed to solve the pickup dispatching problem and the job selection problem simultaneously. In the consider?ation of these studies，a reinforcement learningbased approach was used to select the best dispatch?ing rule for a multiple-load carrier system in a gener?al assembly line［17-18］. Although these methods were sufficiently flexible and efficient，the system dead?lock problem was not addressed yet.

In fact，dispatching rules should be properly de?signed to ensure that a system is deadlock-free.There are two types of deadlock in the AGVs：rout?ing deadlock caused by the path conflict between several AGVs and dispatching deadlock caused by insufficient buffer capacity of the system. The for?mer was usually resolved by means of a zone control traffic management method［19-20］，while the latter should be handled by properly designing the dis?patching methods.

There are two kinds of approaches to handle the deadlock problem：（1）deadlock detection and elimination；（2）deadlock-avoidance. In accordance with some studies［21-23］，deadlock-avoidance is more efficient than the other solution. A resource-orientat?ed colored Petri net was used to model a single-load AGVs［21］，on which a deadlock-avoidance strategy was developed based. However，some studies［22-23］indicate that Petri nets also have disadvantages in terms of complexity and state space explosion. A re?maining buffer capacity-based deadlock-avoidance policy was proposed by Guan and Dai［23］to ensure a deadlock-free system and a high-efficiency dispatch?ing performance simultaneously. Although the simu?lation results proved the effectiveness of the policy，it is only suitable for single-load AGV system.

To the authors’knowledge，the problem of deadlock-avoidance dispatching of multiple-load AGVs is seldom addressed in the pertinent literature.Several researchers［13-16］have tried to reduce the possi?bility of dispatching deadlock by using a traditional strategy，which keeps the number of jobs in the sys?tem（NJIS）at a low level. However，a low level NJIS will make the processing machine easier to idle，so it unavoidably reduces the system efficiency.

The motivation behind this article lies in devel?oping a deadlock-avoidance dispatching method for the multiple-load AGVs. Our main contributions are twofold. On the one hand，a deadlock-avoidance strategy based on the remaining capacity of the workstation is designed to keep the multiple-load AGVs deadlock-free at a higher NJIS level. It can then be incorporated into a multiple-load AGVs dis?patching procedure with four dispatching problems.On the other hand，most of the existing dispatching rules for the dispatching problems are single-attri?bute rules，which only reflect one aspect of the sys?tem. In order to describe multiple aspects of the sys?tem，some system states，e.g. the traveling distance and the buffer capacity，are selected to design sever?al multi-attribute dispatching rules. Finally，several simulation experiments are carried out to evaluate the performance of the proposed deadlock-avoidance strategy and the multi-attribute dispatching rules.

1 Problem Statement

A manufacturing system is composed of a logis?tics transportation sub-system，a processing sub-sys?tem and a set of jobs that need to be processed. The processing sub-system consists of a series of work?stations. The logistics transportation sub-system in?cludes a guide path network and numerous multipleload AGVs. Each job is handled under a set of pre?cedence constraints. Fig.1 gives the layout of a man?ufacturing system as an example. There are fifteen workstations，including an input workstation，an output workstation，and thirteen processing work?stations. Each processing workstation has a process?ing machine，an input buffer and an output buffer with finite capacity. The input workstation has only an output buffer，and the output workstation has on?ly an input buffer. All jobs enter the system via the input workstation. When a job is treated in the man?ufacturing system，it needs to be transported by AGVs and processed by machines. After a series of transport and processing tasks，the job leaves the manufacturing system from the output workstation.Each job has its own task sequences. Generally，the workstation of each processing task is predefined，and this paper focuses on the dispatching of multipleload AGVs.

Fig.1 An example of manufacturing system

First，some important assumptions are given as follows.

（1）The processing machine of each worksta?tion processes the jobs in the input buffer based on a first-come-first-served basis，while jobs in the out?put buffers can be picked up by the multiple-load AGVs in any order in accordance with scheduling re?quirements.

（2）A processing machine processes only one job at a time，and the conversion time for process?ing different types of job is neglected.

（3）All segments of the guide path network are unidirectional.

（4）All AGVs are multiple-load AGVs with the same capacity.

（5）All AGVs travel along the shortest path between pick-up and drop points at a constant speed.

（6）The loading and unloading times for differ?ent types of jobs are the same. When one AGV is loading or unloading at a pick-up or drop point，oth?er AGVs with the similar operations at this point must wait.

Before introducing the improved dispatching procedure，several notations and definitions are giv?en as follows.

W(m)—Workstation m.

CI(m)—Input buffer capacity of W(m).

cI(m)—Current number of jobs in the input buffer of W(m).

CO(m)—Output buffer capacity of W(m).

cO(m)—Current number of jobs in the output buffer of W(m).

P(m)—Number of jobs on the processing ma?chine of W(m)，with P(m) =0 if the processing machine of W(m)is empty；otherwise，P(m) =1.

LO(m)—Number of jobs that have been as?signed to a vehicle，and these jobs still in the output buffer in the output buffer of W(m).

LI(m)—Number of jobs that have been as?signed to a vehicle，and their destination worksta?tion are W(m).

U(m)—Remaining capacity of W(m). It can be calculated by

NR—Number of multiple-load AGVs in the manufacturing system.

Blocked AGV—A multiple-load AGV is called blocked if it has picked up at least one job whose destination workstation has no remaining ca?pacity.

NB—Number of blocked multiple-load AGVs in the manufacturing system.

NBM(m)—Number of blocked multi0070leload AGVs at the drop point of W(m).

Active AGV—A multiple-load AGV that is not blocked.

TA—Task set waiting to be assigned to AGVs in the system，which can be denoted as

TA={Tmj…Tij…Tjm}，where Tijis a task with W( j) as the destination workstation and W(i)as the source workstation.

2 Deadlock?Avoidance Dispatching Procedure

Although Ho et al.［15］proposed a dispatching procedure（DP）for multiple-load AGVs，they did not consider the dispatching deadlock caused by the finite buffer capacity of the workstation. To improve the DP，a deadlock-avoidance dispatching strategy is integrated. The complete flowchart of the im?proved dispatching procedure is shown in Fig.2.

2.1 Deadlock avoidance strategy

To activate a blocked AGV， other active AGVs need to release enough capacity by picking up some jobs in the output buffer of the destination workstation. In this sense，if NB=NR，no active AGV can release new capacity of the output buffer for other blocked AGVs. Consequently，the whole system enters a deadlock situation. To avoid the above-mentioned deadlock situations，it is neces?sary to ensure that there must be at least one active multiple-load AGV in the system. Thus，the dead?lock avoidance strategy can be designed as follows.

（1）Only active multiple-load AGVs can re?ceive new pick-up tasks. According to this rule，one AGV can take at most one job，whose destination workstation has no remaining capacity.

（2）If there is only one active multiple-load AGV in the system，some tasks should be tempo?rarily forbidden. The temporarily forbidden task set is denoted as

Moreover，the available task set TC is given as

Generally，multiple AGVs are used in a sys?tem. So，NR＞1，and the deadlock avoidance prop?erty of the proposed strategy is proven as follows.

Proof 1

If the system has more than one active AGV，all tasks in the system belong to the available task set，and the system is deadlock-free.

Fig.2 Flowchart of the deadlock-avoidance dispatching procedure

Proof 2

If the system has only one active multiple-load AGV，AGVk. For any task，Tij∈TA，according to the values of U( j)and NBM(i)，there are three pos?sible cases：

Case 1 U( j) ＞0 and NBM(i) ≥0

Case 2 U( j) ≤0 and NBM(i) ＞0

Case 3 U( j) ≤0 and NBM(i) =0

In case 1，the destination workstation of task Tijhas enough capacity. If task Tijis assigned to AGVk，then AGVkwill keep active.

In case 2，the destination workstation of task Tijhas no remaining capacity. AGVkwill be blocked if task Tijis assigned to it. However，AGVkwill pick up Tijon W(i) and release a space for blocked AGVs in W(i). As NBM(i) ＞0，the source work?station of task Tijhas blocked at least one multipleload AGV，and one of the blocked AGVs，e.g.，AGVn，will become active. Since the system still has at least one active AGV，it will not become deadlocked.

In case 3，the destination workstation of the task Tijhas no remaining capacity. If the task Tijis assigned to AGVk，then AGVkwill be blocked.Since NBM(i) =0，the source workstation has no blocked AGVs，and no blocked AGV can become active. Then，the system has no active AGV，and will enter the deadlock situation. However，if task Tijis forbidden according to the deadlock avoidance strategy proposed above，it will not be assigned to AGVk，and AGVkwill not be blocked.

Proof 3

As NR＞1，if the system has only one active AGV，there must be at least one workstation has blocked AGVs. Then，all tasks in those worksta?tions belong to case 2. All these tasks will not be forbidden. Thus，the situation in which all tasks are forbidden will not happen.

It is seen from Proofs 1，2 and 3，the deadlock avoidance strategy can keep at least one AGV active in the system and ensure that at least one task is not forbidden. Thus，the system is bound to deadlockfree.

2.2 Dispatching procedure

Based on the DP and the deadlock avoidance strategy proposed above，the basic flowchart of the improved dispatching procedure is shown in Fig.2.The specific operations are described as follows.

Step 1If one AGV（AGVk）has just complet?ed a pickup or delivery task，then considers the sta?tus of the AGV. If it is not empty，go to Step 2；otherwise，go to Step 3.

Step 2Consider the first problem of the dis?patching process—the task selection problem. Dif?ferent task selection rules can be used to determine the next task for AGVk. If the next task is a pickup task，go to Step 4；otherwise，go to Step 11.

Step 3AGVkbecomes idle and waits for new pickup tasks. Once new pickup tasks arise，go to Step 4.

Step 4The deadlock avoidance strategy is used to obtain the available task set TC={Tmj…Tij…Tjm}. Then，the available pickup points set，TPP，can be defined as

Step 5Consider the second problem of the dispatching process—the pickup dispatching prob?lem. The pickup dispatching rule is used to select a pickup point，Pi，from the available pickup point set TPP. Then，dispatch AGVkto the pickup point Pi，and go to Step 6.

Step 6Wait until AGVkarrives at pickup point Pi，and then go to Step 7.

Step 7The deadlock avoidance strategy is used to obtain the available task set TC={Tmj…Tij…Tjm}. Then，the available task set TC(i)of workstation W(i)can be defined as

Step 8If the available task set，TC(i)，is null，go to Step 1；otherwise，go to Step 9.

Step 9Consider the third problem of the dis?patching process—the job selection problem. The job selection rule is used to select one task，Tij，from the available task set. Then，LO(i) and LI( j)are updated accordingly.

Step 10Check the status of AGVk. If it is full or blocked，go to Step 11；otherwise，go to Step 7.

Step 11Consider the fourth problem of the dispatching process—the delivery dispatching prob?lem. The delivery dispatching rule is used to deter?mine the next drop point Dj.

Step 12Wait until AGVkarrives at drop point Dj.

Step 13Check the remaining capacity of workstation W( j). If W( j) does not have enough capacity to hold the job，then go to Step 14；other?wise，go to Step 15.

Step 14AGVkremains stationary until W( j)has enough capacity to accept the job. Then，go to Step 15.

Step 15AGVkdrops the jobs whose destina?tion is W( j). And AGVkbecomes active. Then，go to Step 1.

2.3 Rules adopted for the four problems

Ho et al.［13-15］proposed some relevant rules for the four dispatching problems. These rules can also be adopted by the improved DP. However，most of these rules are single-attribute rules. Since a single attribute cannot reflect all aspects of the system state， we propose the following multi-attribute rules.

2.3.1 The multi?attribute rule for the pickup dispatching problem（MARP）

To reflect various aspects of the system state，the output buffer status and the travelling distance is chosen to calculate the pickup-dispatching utility val?ues. Suppose the available task set is TC={Tmj…Tij…Tjm}and the available pickup point set is TPP={Pi︱?Tij∈TC}. The utility values between AGVkand a pickup point Pican be calculated by

where FD(Pi) and FO(Pi) are the values of the two attributes. AGVkwill visit the pickup point with the maximum utility value.

FD(Pi) is the travelling distance attribute be?tween AGVkand pickup point Pi，and it can be ob?tained by

where MDP=MAX(AGVk，TPP) is the maxi?mum distance between AGVkand any pickup point in the available pickup point set TPP. D(Pi) is the distance between AGVkand pickup point Pi.

The output buffer status attribute FO(Pi) can be defined as

2.3.2 The multi?attribute rule for the job selec?tion problem（MARJ）

Suppose AGVkhas arrived at pickup point Pi.The travelling distance and the input buffer status are chosen as the dispatching attributes of job selec?tion. The available task set of workstations W(i) is TC(i) ={Tij︱?Tij∈TC}. The utility values be?tween AGVkand the job delivering task Tijcan be calculated by

where FD(Tij) and FI(Tij) are the values of the two attributes. AGVkwill pick up the job with the maxi?mum utility value.

FD(Tij) is the travelling distance attribute be?tween the destination of task Tijand the destination of any job currently on AGVk，and it can be calculat?ed by

where MDT is the maximum distance between any pair of drop points in the system. D(Tij) is the mini?mum distance between the destination of task Tijand that of any job currently on AGVk.The input buffer status attribute FI(Tij)is defined as

2.3.3 The multi?attribute rule for the delivery dispatching problem（MARD）

If AGVkhas picked up several jobs and its next task is a delivery task，the multi-attribute rule is used to identify which drop point AGVkshould visit first. Suppose the job set that AGVkhas picked up is TJ={Tmj…Tij…Tjm}. Then，the drop point set can be defined as TPD={Dj︱?Tij∈TJ}. Travelling distance and input buffer status are selected to calcu?late the utility value between AGVkand a drop point in TPD.The utility value is defined as

where FD(Dj) and FI(Dj) are the values of the two attributes. AGVkwill visit the drop point with the maximum utility value first.

FI(Dj) is the travelling distance attribute be?tween the drop point Djand the current position of AGVk，which can be calculated by

where MDD=MAX(AGVk，TPD) is the maxi?mum distance between AGVkand all drop points in the drop point set TPD. D(Dj) is the distance be?tween AGVkand the drop point Dj.

The input buffer status attribute FI(Dj) is de?fined as

3 Simulation Experiments

3.1 Simulation model

To verify the performance of the deadlock avoidance strategy and the multi-attribute rules pro?posed in this paper，simulation tests are carried out to compare the TS with our strategy under eight combinations of rules. The rules adopted in our sim?ulation are introduced as follows.

3.1.1 Rules adopted for the task selection prob?lem

Ho et al.［14］proposed three single-attribute rules for this problem and compared their perfor?mance results. Simulation results showed that the DTF rules had the best throughput performance.Hence，the task selection rule adopted in our simula?tion experiments is DTF. According to this rule，if the AGV is partially loaded，it will always perform a delivery task.

3.1.2 Rules adopted for the pickup dispatching problem

Two rules are adopted in our simulation experi?ments. One is the proposed MARP，and the other is the greatest queue length（GQL）rule. According to Ref.［13］，GQL is the best among the twelve pickup dispatching rules proposed in their paper.Based on the GQL rule，one AGV（AGVk）will visit the pickup point that has the greatest number of jobs waiting at its output buffer.

3.1.3 Rules adopted for the job selection prob?lem

The first rule adopted in our simulation experi?ments is MARJ. Additionally，Ho and Liu［15］stud?ied six job selection rules to evaluate their perfor?mances. Their work has showed that the identical destination first（IDF） rule had the best perfor?mance. Hence，the IDF rule is adopted as the sec?ond rule. When the IDF rule is used，one AGV（AGVk）will pick up the job whose destination is the same as the next destination of any load current?ly on AGVk.

3.1.4 Rules adopted for the delivery dispatch?ing problem

The first rule we adopted is MARD. The sec?ond is the shortest distance（SD） rule. Ho and Chien［14］have proved that the SD rule is better than other delivery dispatching rules in all performance measures. This rule ensures that the job with the smallest distance between its destination and the cur?rent position of AGV will be delivered first.

The eight hybrid-rule sets are described in Ta?ble 1. An example of a manufacturing system is shown in Fig.1. The capacity of the buffer in the in?put workstation is 4，and that of the input and out?put buffers in every processing workstation is 2.The processing information of the job set is listed in Table 2. The processing time for every job in every workstation is 90 s.

Table 1 All hybrid?rule sets adopted in this study

Table 2 The processing information of the job set

There are ten multiple-load AGVs in the sys?tem. Each AGV has a 4-item loading capacity and travels at the speed of 1 m/s. It takes 30 s for one AGV to perform a loading or unloading operation.The software Tecnomatix Plant Simulation 15.0 is utilized in our simulation tests，with a software in?terface as shown in Fig.3.

Fig.3 Simulation experiment interface

3.2 Tests using the traditional strategy

First，the traditional strategy is used in the sim?ulation tests. The simulation duration is 96 h，and each kind of test repeats 10 times. The deadlock times for the traditional strategy with different hy?brid-rule sets are shown in Fig.4. and the average throughput for this strategy is shown in Fig.5.

Fig.4 Deadlock times for the traditional strategy under all hybrid-rule sets

Moreover，the simulation results show that the hybrid-rule set with code 1222 enters deadlock most probably for all NJIS levels. When the NJIS level is 55，this hybrid-rule set enters deadlock very early and produces no jobs. Moreover， the average throughput of the hybrid-rule set with code 1111 is 0.8%—64.4% higher than the other hybrid-rule sets. Based on the obtained results，it is obvious that the multi-attribute rules can synthesize the mul?tiple aspects of the system.

Fig.5 Average throughput for the traditional strategy under all hybrid-rule sets

3.3 Tests using the deadlock avoidance strategy

In this subsection，simulation tests are conduct?ed to validate the deadlock-avoidance strategy pro?posed in this paper. The system parameter settings are the same as the previous one. When the dead?lock-avoidance strategy is used，no deadlock occurs for any hybrid-rule set at any NJIS level. The aver?age throughput for the deadlock-avoidance strategy with different hybrid-rule sets is shown in Fig.6.

Fig.6 Average throughput for the deadlock avoidance strat?egy under all hybrid-rule sets

In Fig.6，when the deadlock avoidance strate?gy is used，the best NJIS levels for all hybrid-rule sets are more than 50. In contrast，if the traditional strategy is used，all hybrid-rule sets cannot prevent a deadlock. It is obvious that the deadlock-avoid?ance strategy is indispensable. The comparison of Figs.5 and 6 shows that，when the deadlock-avoid?ance strategy is adopted，the average throughput for all hybrid-rule sets at each NJIS level increase.Take the hybrid-rule set with code 1121 for exam?ple，when using the proposed deadlock-avoidance strategy，the optimal average throughput is 729.4，which is 5.2% higher than 693.1 with the traditional strategy. This result implies that the proposed dead?lock-avoidance strategy can make full use of system resources and improve the performance of dispatch?ing rules.

As shown in Fig.6，the hybrid-rule set with code 1111 performs better than any other set at each NJIS level. The optimal average throughput of the hybrid-rule set with code 1111 is 730.5，which is 0.2%—9.7% higher than the other hybrid-rule sets.This result indicates that the multi-attribute rules are adaptive to various system settings and more ef?ficient than single-attribute rules for multiple-load AGVs with finite buffer capacity constraints.

Fig.7 Average system throughput, working rate, blocking rate and starvation rate of the processing machines under the hybrid-rule set with code 1221

The simulation results also show that the NJIS level has a significant effect on the system through?put. Take the hybrid-rule set with code 1221 for ex?ample，the system throughput，working rate，block?ing rate and starvation rate of the processing ma?chines are shown in Fig.7. The throughput perfor?mance is directly proportional to the average work?ing rate of the processing machines. If the NJIS lev?el is too low，the processing machines are frequent?ly starved. On the other hand，if the NJIS level is too high，the processing machines are likely to be?come blocked as the output buffer is frequently full.Both of these two situations will reduce the average working rate of the processing machines；thus，the efficiency of the processing sub-system being de?creased.It is obvious that the NJIS level has a signif?icant effect on the system throughput. Every hybridrule set has a best NJIS level. When operated below or above this level，the system efficiency will be re?duced.

4 Conclusions

The dispatching problem of multiple-load AGVs with finite buffer capacity constraints is pre?sented. The deadlock-avoidance dispatching strate?gy based on the remaining capacity of the worksta?tion is integrated with a multiple-load AGV dis?patching procedure. The following findings can be obtained from the simulation results.

A proper NJIS level is critical to manufacturing systems with finite buffer capacity constraints. Al?though a lower NJIS level can prevent the system from reaching deadlock，this approach will also re?duce the system efficiency. The dispatching problem of multiple-load AGVs with finite buffer capacity constraints is addressed for a manufacturing system operated at a high NJIS level. In order to improve the system efficiency at a higher NJIS level while still avoiding deadlock，the deadlock-avoidance dis?patching strategy based on the remaining capacity of the workstation is integrated with a multiple-load AGV dispatching procedure. Simulation results show that the proposed deadlock-avoidance strategy can make full use of system resources and improve the performance of dispatching rules. Moreover，the simulation results also show that the optimal aver?age throughput of the hybrid-rule set with code 1111 is 0.2%—9.7% higher than the other hybrid-rule sets. This indicates that the proposed multi-attribute rules are adaptive to various system settings and more efficient than single-attribute rules for multipleload AGVs with finite buffer capacity constraints. It is hopeful that these research findings can be helpful for solving multiple-load AGV dispatching prob?lems in similar industrial scenarios. However，the intrinsic relationship of the NJIS level with the buf?fer size of the workstations or the multiple-load AGVs is still not fully clarified yet. Hence，it will be investigated further in our future research work.